17 Nov 2026 · 10:44 AM MTUpdated…
Build plan

From idea to evidence.

Every commitment, its owner, its next step and its recorded evidence. Search by subject or item number.

Research decisions ↗
Where the build stands

290 recorded items marked done; 0 in progress; 15 queued. These board states do not establish a measured model gain.

Machine map and recorded totals

The machine map data 2m ago

99%
RECORDED PLAN WORK2 open filings

134 of 136 marked done · counted once across the machine

The ring counts historical board state.
It does not certify requirements or the live product.

Requirement acceptance: 0 of 3 assessed accepted; 133 unassessed

Recorded Landed work is separate from accepted, deployed and operating evidence.

E-01 · implementation recorded · test negative · acceptance pending · deployment unverified · operation unverified. Next: Complete the prescribed rolling origins and meet the held-out quality gate on the exact release.

L-10 · implementation recorded · test partial · acceptance pending · deployment unverified · operation unverified. Next: The exact promotion path must reject a planted paired whole-game regression and bind a qualifying candidate to the served release.

N-04 · implementation recorded · test rehearsal · acceptance pending · deployment unverified · operation unverified. Next: Verify actual September 29 TUE-lock issuance: exact release, all games, WORLD rows, production simulation count, $0 proving rows and immutable replay.

01 / RESEARCH & BUILD02 / ONE FOOTBALL WORLD03 / OPERATIONS, DECISIONS & LEARNINGExternal researchAMBER · Human-run intakeManual research workflowResearch questionsGREEN · Work marked done5/5 done · 0 leftBuild & challengeGREEN · Roles scheduled5/5 roles loadedEvaluate changesGREEN · Work marked done10/10 done · 0 leftModel capacityRED · Allowance LOWClaude 32% · Codex 16% · Antigravity · Gemini models 0% · Antigravity · Claude and GPT models 0%Shared work & evidenceAMBER · Plan completion134/136 done · 2 leftSource dataGREEN · Work marked done16/16 done · 0 leftWarehouseGREEN · Work marked done6/6 done · 0 leftS · StatisticsGREEN · Work marked done16/16 done · 0 leftB · BehaviorGREEN · Work marked done7/7 done · 0 leftQ · QualitativeGREEN · Work marked done4/4 done · 0 leftNFL world modelGREEN · Work marked done19/19 done · 0 leftStored productsAMBER · Needs attention10/10 done · 0 leftYour football siteGREEN · Checks pass6/6 done · 0 leftKeep it runningGREEN · Checks pass6/6 done · 0 leftNew experimentsAMBER · Experiments registered6 registered · results dueReality & scoringGREEN · Work marked done10/10 done · 0 leftPrediction locksUNKNOWN · Coverage not measured5 horizons · coverage unknownBetting & fantasyRED · Staking gate CLOSED6/6 done · 0 leftOUTCOMES → ERRORS → NEW QUESTIONS → BETTER MODEL
GREEN · recorded done / checks passAMBER · partial / attentionRED · not started / blocked / low allowanceUNKNOWN · evidence missing

Arrows show the architecture and intended handoffs, not live traffic. Plan completion is separate from runtime health and prediction accuracy. Snapshot Nov 17 · 10:41 AM MT. Select a component for its evidence.

Recorded items305

Includes project operations

Marked done290

Board state; evidence below

In progress0

See Who is working for live jobs

Queued15

Ordered by recorded priority

A-03A-03 — Bets go down early, so forecast the injury report itself: P(Friday status \| Tuesday informationdone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item A-03, lane ratings, lineups, policies, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Bets go down early, so forecast the injury report itself: P(Friday status \| Tuesday information). | Brier against the base rate by body part. FILES THE PLAN NAMES: research/playgrain/lineup/inj_report_forecast.py MISSING: inj_report_forecast.py No prior board item ever mentioned this.

A-04A-04 — Per-snap exit hazard by position, age, injury history, surface and workload. Inside the simulatodone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item A-04, lane ratings, lineups, policies, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Per-snap exit hazard by position, age, injury history, surface and workload. Inside the simulator a player can leave and his backup enters. | Simulated exits per game match history by position. FILES THE PLAN NAMES: research/playgrain/lineup/ingame_injury_hazard.py MISSING: ingame_injury_hazard.py No prior board item ever mentioned this.

E-01E-01 — Rolling origin 2004→2025: fit ≤ season s−2, calibrate on s−1, simulate every game of s from its done

Owner: Codex · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item E-01, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Rolling origin 2004→2025: fit ≤ season s−2, calibrate on s−1, simulate every game of s from its as-of state. Produces a distribution for every game in 22 seasons. Runs overnight; resumable. FILES THE PLAN NAMES: research/playgrain/evalx/walkforward.py MISSING: walkforward.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10. HOW TO RUN IT IN A UNIT, added 2026-09-22 after three units were killed at the alarm and the lane was restarted mid-origin: `--limit 1`. One origin is ~9 min (2008: fit 130s + sim 412s); two fit a 20-minute unit, three do not. Origin 2008 is checkpointed in data/playgrain-walkforward/manifest.json; 2009 was lost to the restart and is the resume point. Run one or two origins, exit clean, let the next unit continue. REOPENED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): FALSE DONE. All 18 origins ran 09-22 13:41 to 09-23 04:42, before the S-08 magnitude-head fix c51817b2 (09-23 16:32): mean simulated total 81.6-89.3 vs actual 45-50 (2014/2020/2024/2025), margin-result correlation 0.04-0.22. data/playgrain-walkforward/preds/*.csv are INVALID; preds-current-kernel/2025.csv still totals 88.5 because it reused the stale models/. RE-RUN ON THE S-16 RELEASE: (1) add --out DIR to walkforward.py (manifest, models, preds under data/playgrain-walkforward/<release_id>/) and write INVALID.md into the old preds/ citing c51817b2 -- keep the old run as evidence, never overwrite it; (2) origins 2021-2025 first (the early-line seasons), then 2008-2020, n=300, refitting every origin with the post-fix fit_models.py; (3) have it write a progress json {done:[origins], total:18} and register a `grind` field on this item so bin/grind.py runs the rest at zero model tokens: cmd '.venv/bin/python research/playgrain/evalx/walkforward.py --out data/playgrain-walkforward/<release_id> --start 2021 --limit 1 --n 300' (then --start 2008), state_file = that progress json, state_key 'done', total 18, per_run_s 1800. ACCEPTANCE per origin: mean simulated total within 1.5 pts of actual and margin-result correlation >= 0.30 (the 2024 held-out kernel reached 0.418); otherwise stop and report. S-16 kernel release frozen 2026-09-24 22:25 MDT: <release_id> = k51f2563814f06fca7f79 (data/sim_store/releases/k51f2563814f06fca7f79/, listed in data/sim_store/releases.json). Load with research.playgrain.kernel_release.load_sim(id, seed) or Sim(models_dir=<that dir>); cite it on every walk-forward origin. GEMINI 1/10 (2026-09-25, compounding error): per origin also report sim vs league per drive and per quarter — points/drive, plays/drive, penalty yards/game, 3rd-down rate — so drift is caught where it starts, before any top-down regulariser is considered. E-01-QB FINDING (2026-09-25 02:02 MT, Codex; confirmed claude): walkforward.py passed qb=0.0 for BOTH teams on every historical game, while the Tuesday issuer injects each side's as-of QB rating and the heads were trained with qb_epa. Every k51f E-01 acceptance / S-06 FAIL / E-03 G4 RED / 1-2pt edge verdict measured a model WITHOUT quarterbacks, not the product. Fix 4ce67707: walkforward --qb {none,prior-starter,actual-starter}, separate dirs data/playgrain-walkforward/k51f2563814f06fca7f79-qb-<mode>/, models reused read-only; scored by evalx/e01_qb_compare.py. RESULTS 2026-09-25 04:50 MT (2024+2025 prior-starter vs no-QB, same 544 early-priced games, raw, nothing tuned): margin RMSE paired -0.32 [-0.53,-0.11]; corr .423/.391 -> .429/.409; slope 1.82 -> 1.29 [1.06,1.51]; pooled early return -3.83% [-9.75,+2.15] -> -1.32% [-7.06,+4.46]; spread CLV -0.087 -> +0.017pt; G4 still RED; predeclared 1-2pt rule still FAILS. Totals +1.90/+1.97 -> both QB origins REJECTED on the total test; release drift_shift.json was ALSO calibrated on qb=0 sims (calibrate_drift.py:27-28). Not run: actual-starter, 2021/2023. Full: data/coord/e01-qb-result.md, evalx/e01_qb_compare.k51f2563814f06fca7f79.json (commits 4ce67707, a9c0eeeb). HANDOFF 2026-09-25 10:35 MT (Claude): origins 2008-2020 are running as 3 detached walkforward.py processes (--continue-past-rejected, logs data/logs/e01-sweep-{1,2,3}.log, reniced +15). Do NOT start more. CLOSE when data/playgrain-walkforward/k51f2563814f06fca7f79/progress.json has done+rejected = 18: state done, result = the per-origin table (total_err_pts, margin_result_corr) from manifest.json. Pattern so far: 2021/2023-2025 pass; 2008-2016 and 2022 rejected on totals (sim 2-3 pts low in older eras) because the frozen release applies today's scoring constants to every era. That is a measured result, not a reason to hold the item open.

F-04F-04 — Move the 0.40 FantasyPros blend, league-rate calibration, designation decay and rookie pass *insdone

Owner: Codex · Lane: truth and the fantasy path · Priority: 0

Why this matters

Plan item F-04, lane truth and the fantasy path, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Move the 0.40 FantasyPros blend, league-rate calibration, designation decay and rookie pass *inside* `lines_for`, or score the board end to end, so the board and the measured model are one object. | Live mean error −3.42 → ~0 FILES THE PLAN NAMES: none all present No prior board item ever mentioned this.

G-00G-00 — Game-model recipe type and served-pointer switch-over in the release loop (team-strength project, first unit)done

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

docs/research/TEAM-STRENGTH-PROJECT.md milestone 4 needs a game-model candidate to reach picks, pick'em and the site automatically, and it cannot today. Checked 2026-10-06: playerweek/auto_release.py recognises exactly two recipe schemas, release-candidate/1 (player model) and kernel-candidate/1 (Track B world-model kernel, RESEARCH-CHARTER.md §14, wired by commit 16f4a40f9). Neither is a game-model (team-layer margin/total) recipe. bin/freeze-games.py:289 calls `teamstate.predict(games, teamstate.PUBLISHED)` directly — a hardcoded module constant, not a release pointer — so there is no mechanism by which a better team-strength candidate could ever become served. The shared game-model harness this recipe gates on already exists and is approved: research/scientist/harness/game_pit.py (built by commit ec95084e4, RESEARCH-CHARTER.md §6a) plus research/betting/money_backtest.py and bin/headline-metrics.py (playerweek/exprecord.py APPROVED_HARNESSES). Only the release-loop wiring is missing.

Evidence and next step

Add schema "game-candidate/1" to playerweek/auto_release.py, modelled on the existing kernel-candidate/1 path (BUILD/GATE/PROMOTE/WATCH, same module doc). GATE on research/scientist/harness/game_pit.py (or research/betting/money_backtest.py + bin/headline-metrics.py while game_pit integration is new): paired winner % vs Vegas, margin MAE and total MAE on the same games, computed overall AND for weeks 10-18 separately, non-inferior or better than the served game model with the interval not showing harm — the same discipline the player-model gate already applies (NONINF_PTS). On PROMOTE, write data/releases/served-game-model.json atomically (parallel to the existing data/releases/served-release.json) naming the promoted candidate. Change bin/freeze-games.py:289 to resolve that pointer once per run (the existing "resolve once at start, pass explicitly through the run" rule in CLAUDE.md §"Releases ship when they pass") and fall back to teamstate.PUBLISHED when no candidate has ever been promoted, so nothing already served breaks. Because picks (data/proving.jsonl), pick'em and the site all read what freeze-games.py writes, this one change is the whole switch-over; no separate pick'em/site edit is expected unless one of them is found to call teamstate.predict directly (grep before editing — research/preseason_edge.py, playerweek/season_sim.py and several research/bin scripts call it too but are not the served path and are out of this unit's scope). ACCEPTANCE: a drill candidate ("test": true) carries a game-candidate/1 recipe through build, gate and promote without touching the real pointer; a planted-regression drill candidate (margin MAE deliberately worse) is refused by the gate; bin/freeze-games.py reads data/releases/served-game-model.json when it exists and teamstate.PUBLISHED when it does not (both paths covered by a planted test); the one-line coordination note this unit requires is already posted (data/coord/claude-codex.md, 2026-10-06).

G-01G-01 — Team-strength milestone 1, unit 1 of 4: point-in-time efficiency data assemblydone

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

CLAUDE.md 2026-10-06 "THE FOCUS": we pick 64.1% of winners against Vegas's 66.3%; the ceiling study found the gap lives in team strength. H-377 (CLAIMED): D.ppd carries 89% of the margin residual and EPA/play 80% (research/scientist/RESIDUAL-MAP.md); the gap grew to 6.3, 4.4, 2.9 winners/100 in 2023-25. docs/research/TEAM-STRENGTH-PROJECT.md milestone 1 names the inputs: EPA per play, success rate, points per drive, explosives, turnovers treated as mostly noise, separate pass and run, 2016-2026.

Evidence and next step

Build research/teamstrength/assemble_efficiency.py: one point-in-time table keyed (team, season, week, side[off/def], unit[pass/run]) with EPA/play, success rate, points per drive (from the existing drive-level tables), explosive-play rate and a heavily down-weighted turnover rate, from the play-level data already in the warehouse (`plays`, `plays_wide`, P-06). Every row for week w must read only plays from games that kicked off strictly before week w's own games — the exact as-of discipline PW-143 found broken in team_latents (an open incident, INC-616354ef39); this new table must not repeat that leak. Use bin/auto-release-arm.py's AsOf view or an equivalent explicit as-of filter, never a bare SELECT against `plays`. ACCEPTANCE: one row per (team, season>=2016, week, side, unit) with no gaps; a planted-future control (handing the builder next week's own plays) is refused by a new test (tests/test_team_efficiency_pit_asof.py) mirroring H-377's own card control ("hand the model next week's EPA and the harness must refuse").

G-02G-02 — Team-strength milestone 1, unit 2 of 4: opponent-adjusted rating fitdone

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

playerweek/teamstate.py's own docstring: "S evidence only: completed scores, identities, season/week and fixtures" — v0.1 learns only from final scores. docs/research/TEAM-STRENGTH-PROJECT.md milestone 1: opponent-adjusted offense/defense ratings from play-level data, separate pass and run, shrunk toward the prior. Depends on G-01's table.

Evidence and next step

Build research/teamstrength/ratings.py: fit opponent-adjusted offense and defense efficiency ratings from G-01's table, separately for pass and run, with EPA/play and success rate as the primary signal, points-per-drive as the scoring bridge, explosives secondary, turnovers shrunk hard toward the league mean. Shrink every rating toward its prior by games-seen (same discipline as T-09's shrinkage audit). Expose `predict(asof_con, season, week) -> rows(game_id, margin, total, p_home)` with the exact signature research/scientist/harness/game_pit.py's `score()` expects, so it can be registered as a candidate beside fc_v01 in research/betting/money_backtest.py with no new scaffolding. ACCEPTANCE: research/scientist/harness/game_pit.py's three leak checks (score_truncation, line_shift, planted_future) all pass on this candidate; pass-vs-run and offense-vs-defense coefficients are fit and reported separately, not pooled.

G-03G-03 — Team-strength milestone 1, unit 3 of 4: walk-forward evaluation on the shared harnessdone

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

docs/research/TEAM-STRENGTH-PROJECT.md: "Standalone test: winner %, margin and total error vs v0.1 and vs Vegas, by season stage." CLAUDE.md: "Every milestone records winner % vs Vegas on the same games, overall and for weeks 10-18, with n and an interval." Depends on G-02's candidate.

Evidence and next step

Run G-02's candidate through research/scientist/harness/game_pit.py (approved harness, RESEARCH-CHARTER.md §6a): discovery 2016-2020, replication 2021-2022, holdout 2023-2025 (H-377's own card split). Report paired winner % vs Vegas (discordant-pair comparison, RESEARCH-CHARTER.md §6), margin MAE and total MAE against both v0.1 (teamstate.PUBLISHED) and the Vegas closing line, computed overall AND for weeks 10-18 separately — the 3.4-point gap CLAUDE.md names — never blended into the overall number. Game-block bootstrap intervals over (season, week) slates. ACCEPTANCE: a results table with n and 95% CI for every (population × model × metric) cell, with the weeks 10-18 slice reported as its own row, written to research/scientist/experiments/.

G-04G-04 — Team-strength milestone 1, unit 4 of 4: experiment record and value claim (H-377)done

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

RESEARCH-CHARTER.md §8-9: a result counts only after replication and holdout, written up in standard form with a value_claim block; credit is computed by research/management/credit.py, not by opinion. H-377 is currently CLAIMED with no scored candidate on record. Depends on G-03's measurement.

Evidence and next step

Write the standard-form record (docs/research/EXPERIMENT-STANDARD.md) for H-377 from G-03's results: summary, mechanism (final-score-only state vs play-level efficiency), the self-attack list (RESEARCH-CHARTER.md §7: benchmark, leakage, metric, multiple comparisons, confounds, regime change), and a `value_claim` block (node D.ppd / T.points, node_effect, top_effect with winner_pp / margin_mae / total_mae, discovery/replication/holdout periods, harness="research/scientist/harness/game_pit.py") exactly in the §8 schema. If it beats v0.1 on the shared harness, submit it as a game-model candidate through G-00's game-candidate/1 recipe (once G-00 lands) or record it pending that recipe per RESEARCH-CHARTER.md §14's "recorded but not served" fallback — never hand-edit the served pointer. ACCEPTANCE: one experiment record under research/experiments/ moving from status preregistered to graded with a value_claim block; H-377's row in research/scientist/HYPOTHESES.md updated from CLAIMED to its graded verdict.

G-05G-05 — Team-strength milestone 2: adaptive learning rate (H-378)done

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

H-378 (OPEN): weeks 15-18 lose 6.26 winners/100 to Vegas (n=463) against 0.2 in weeks 1-4 — the largest situational residual on the map. docs/research/TEAM-STRENGTH-PROJECT.md milestone 2: let ratings move at a pace depending on season stage and evidence accumulated; handle teams resting starters in weeks 17-18.

Evidence and next step

Build research/teamstrength/adaptive_learning.py on top of G-04's graded, unpromoted efficiency ratings: a state-space model with time-varying process noise (or a learned schedule) so the update rate depends on games-seen and season stage, plus clinch/elimination state (computed as-of the Tuesday before, from standings) as a pre-kickoff input for rested starters in weeks 17-18. ACCEPTANCE: margin MAE on weeks 15-18 improves with a bootstrap CI excluding 0, or is reported not-worse with the CI shown; winner % vs Vegas reported for weeks 10-18 on the same walk-forward harness as G-03.

G-06G-06 — Team-strength milestone 3: starter-specific QB ratingdone

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

docs/research/TEAM-STRENGTH-PROJECT.md milestone 3: rate the team's passing offense by the expected starter, with a replacement-level prior for low-history QBs, so backup-QB games stop costing us — totals currently run 2.45 points high in QB-change games.

Evidence and next step

Build research/teamstrength/starter_qb.py: replace the team-level passing-offense rating with a starter-conditional one, with a replacement-level prior shrinking in for QBs with little history (same family as T-04's draft-pick / age-curve priors). ACCEPTANCE: total MAE in QB-change games improves vs v0.1 with a reported CI; the replacement-level prior is calibrated against low-history QBs' actual subsequent performance, not asserted.

G-07G-07 — Team-strength milestone 4: team-layer release (H-391)next

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

docs/research/TEAM-STRENGTH-PROJECT.md milestone 4: margin and total come from milestones 1-3, reconciled with v0.1 by accuracy-weighted reconciliation; submit as a game-model candidate to the release loop; picks, pick'em and the site switch to it automatically if it passes. H-391 (OPEN): "the served game numbers (v0.1, from scores) and the player-model roll-up disagree" — this unit is the game-level half of that reconciliation.

Dependency

G-00, G-04, G-05, G-06 (needs the recipe type and milestones 1-3's ratings)

Evidence and next step

Build research/teamstrength/reconcile.py: accuracy-weighted reconciliation of G-04/G-05/G-06's team-layer margin and total with v0.1's, per H-391's design (grade on the shared harness at the node and the top of the tree). Submit the result as a real (non-drill) game-candidate/1 recipe (data/releases/candidates/<id>.json) through G-00's new release-loop path. ACCEPTANCE: the recipe is queued, gated on the shared harness overall and for weeks 10-18, and either promoted (picks/pick'em/site switch over through G-00's pointer with no further edit) or refused with the reason logged in data/releases/auto-release-log.jsonl.

G-08G-08 — Team-strength milestone 5: coherence between team and player layersnext

Owner: — · Lane: ratings, lineups, policies · Priority: 0

Why this matters

docs/research/TEAM-STRENGTH-PROJECT.md milestone 5: the player layer allocates the team layer's volumes, so one top-to-bottom forecast is published (docs/research/STATS-MODEL-PROGRAM.md). H-391's own example: IND@WAS showed different "ours" numbers at the game and player levels.

Dependency

G-07 (needs the promoted team-layer release)

Evidence and next step

Once G-07 ships, make the player layer allocate plays/drives/points from the promoted team-layer forecast instead of carrying its own independent total, per docs/research/STATS-MODEL-PROGRAM.md's single top-to-bottom forecast. ACCEPTANCE: on a sampled slate, the sum of published player stat lines equals the served team total exactly — the same one-ledger invariant S-14 already requires inside the simulator, applied here to the served Stats-model product.

L-10L-10 — The evaluator today scores only the action model. Generalise it to every kernel component (play-done

Owner: Codex · Lane: unassigned · Priority: 0

Why this matters

Plan item L-10, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The evaluator today scores only the action model. Generalise it to every kernel component (play-level log loss, n≈911k) and to a simulator-level paired test under common random numbers. A proposal that improves a component and worsens game CRPS is refused: the aggregation trap is how the share model died. FILES THE PLAN NAMES: research/playgrain/loop/multi_target_evaluator.py MISSING: multi_target_evaluator.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

N-04N-04 — Tuesday issuance: price every game at the TUE lock with the world model (week 4 on 09-29, week 5 on 10-06)done

Owner: Codex · Lane: kernel, drift, decision · Priority: 0

Why this matters

Critical-path step 6 (build audit /tmp/claude-501/build-audit-2026-09-24.md §1, §5). O-06's playgrain-issue job was never built and E-08 issues only teamstate-family models, so the world model has never priced a full slate at a lock. TNF alone had seven different GB margins across records; there is no single frozen game price per lock.

Evidence and next step

INPUTS: the S-16 frozen kernel release id; data/playgrain.duckdb current through the prior week (N-10); research/playgrain/sim.py Sim.play_slate (full game incl. overtime, sim.py:791); research/playgrain/queries.py / serving_adapter.py game products (S-15: margin, total, moneyline); early_lines (C-04) for the book number at issue. BUILD: research/playgrain/evalx/issue_playgrain.py --season 2026 --week W [--rehearse]. For every scheduled game of week W not yet kicked off: n>=2000 trajectories from the as-of state (warehouse.asof(<Tuesday date>)), frozen with sim_store.issue(game_id, Lock TUE, release) (refuses to overwrite); a 4th version row 'WORLD' in playerweek.shadow `forecasts` beside v0.1/POWER/DYN (E-08's table, frozen-means-frozen); fair spread/total/moneyline via queries.py into data/playgrain-issue/2026_wWW.json with release id, sim_store hash per game and issued_at UTC; one $0 row per game in data/proving.jsonl (the model's side vs the early number, stake 0, marked shadow). Schedule bin/com.seano.playerweek.playgrain-issue.plist Tuesdays 07:00 MT (after N-10), LOADED. REHEARSAL: run --week 3 --rehearse on week 3's Sunday/Monday slate before Sun 09-27 11:00 MT; a game already kicked off is refused, never priced. ACCEPTANCE: by Tue 09-29 09:00 MT, `forecasts` version WORLD has one row per week-4 game (count = schedule count); every row cites one release id and a sim_store hash that reloads bit-for-bit; a second run changes nothing; proving.jsonl has one $0 row per game. Same for week 5 by Tue 10-06 09:00 MT. Added to the plan 2026-09-24 (revision 6 of the plan document). Family N: what the objective needs that the design did not name.

N-10N-10 — In-season play-grain refresh: load 2026 weeks 2–3 into data/playgrain.duckdb and keep it current weeklydone

Owner: Codex · Lane: sensors · Priority: 0

Why this matters

Critical-path step 1 (build audit /tmp/claude-501/build-audit-2026-09-24.md §1, §5). The shared data/playgrain.duckdb stops at 2026 week 1: `plays` 2026 = week 1 only (2,756 rows), `team_latents` stops at week 1, `plays_wide` weeks 1–2. Nothing loads it: O-06's playgrain-refresh job was never built (O-06 falsely marked done). Every week-4/5 price is made from a stale as-of state until this lands.

Evidence and next step

INPUTS: data/cache/play_by_play_2026.parquet (kept fresh by C-11's research/playgrain/sensors/nflverse_refresh.py); the existing builders research/playgrain/load_plays.py (`plays`), research/playgrain/load_plays_wide.py (`plays_wide`), research/playgrain/latents.py (`team_latents`, `qb_latents`). BUILD: research/playgrain/sensors/playgrain_refresh.py -- one idempotent chain: nflverse_refresh (pbp 2026) -> load_plays -> load_plays_wide -> latents, every write through the existing write-lock retry discipline (deadline connect, capture yields; see PW-93). One JSON line per run to data/playgrain_refresh.jsonl: started/finished UTC, rc, and for season 2026 max(week) and row count in each of the four tables. Wrap it in bin/playgrain-refresh.sh + bin/com.seano.playerweek.playgrain-refresh.plist: Tuesday 04:00 MT plus a daily 06:00 MT catch-up; LOAD it (launchctl list must show it) and kickstart it once. ACCEPTANCE: (1) tonight: `SELECT max(week) FROM plays WHERE season=2026` = 2, and the same max week in plays_wide, and team_latents/qb_latents carry the as-of row for week 3; (2) by Tue 09-29 06:00 MT (after Monday night) week 3 is loaded in all four tables; (3) per-season row counts for 1999-2025 are unchanged before vs after; (4) an immediate rerun is a no-op. Then note on O-06 that playgrain-refresh is built. N-04 (Tuesday issuance) reads this. Added to the plan 2026-09-24 (revision 6 of the plan document). Family N: what the objective needs that the design did not name.

N-12N-12 — Price every remaining week each Tuesday; accuracy by lead timedone

Owner: Codex · Lane: kernel, drift, decision · Priority: 0

Why this matters

Sean, 2026-09-25 00:55 MT, verbatim: "I said I want all weeks prices each week. i want to know the distribution of accuracy based on weeks/days out." N-04 prices only the coming week at the TUE lock, so no game is ever priced more than ~5 days out and there is no measurement of how accuracy decays with lead time.

Evidence and next step

FORWARD: every Tuesday the N-04 job (bin/playgrain-issue.sh, launchd Tue 07:00 MT) prices the coming week exactly as before (TUE key, WORLD forecasts, proving rows) and then EVERY remaining regular-season week from the same as-of inputs, each (game, issue date) frozen as its own sim_store issuance under a lead-aware key (TUE lock + as-of date), refuse-to-overwrite kept, freshness gate kept. Far-week prices never enter forecasts or proving.jsonl (those stay one per game for the next-week price). HISTORICAL: research/playgrain/evalx/lead_curve.py simulates 2024/2025 (then 2023/2021) regular-season games from the as-of state at leads of 1, 2, 4, 8 weeks with the frozen S-16 release and E-01's origin models (fit <= s-2), n=300, and grades each lead with market_diag conventions: margin MAE and its distribution, CRPS, P(home) log-loss/Brier, calibration slope, ATS / O/U hit and flat -110 return vs the odds-historical ladder line at the same lead and vs the close, ML vs close; game-block bootstrap 95% intervals. Output lead_curve.json + LEAD_CURVE.md. ACCEPTANCE: (1) a rehearsal of 2026 weeks 4-18 from one Tuesday prices every scheduled game into the rehearsal store, each reloadable bit-for-bit, and a rerun changes nothing; (2) existing TUE issuances and their receipts verify unchanged (tests/test_sim_store.py passes); (3) production-n runtime measured and the Tuesday run fits 07:00-12:00 MT (n per lead chosen if not); (4) lead_curve.json has every lead for 2024 and 2025 with intervals; (5) the Tuesday 09-29 run leaves week-5..18 issuances with as-of 2026-09-29 in data/sim_store. Added to the plan 2026-09-25 (revision 7). Family N: what the objective needs that the design did not name.

PW-143As-of leak: team_latents centre each play's EPA on the FULL season average (future games included)next

Owner: Codex · Lane: unassigned · Priority: 0

Why this matters

latents.py:team_game_epa centres EPA by avg(epa) over (partition by season, play_type) on the whole historical scrimmage table, so a week-3 latent uses the season mean through week 18. Every backtest built on these latents (E-01, E-03) carries this look-ahead.

Dependency

Source-admissible multi-origin E-01/E-03 evaluation needs historical source observation clocks, which are unknown; the first-96 2024 replay is complete (paired RMSE diff +0.020, 95% CI [-0.196,+0.218]) and nothing further is runnable.

Evidence and next step

Replace with an expanding as-of mean (prior weeks only, prior-season carry for early weeks), rebuild latents as a new versioned table, and measure how much E-01/E-03 move. Found by Codex during T-09; Codex owns the diagnosis, a new kernel release is required for any fix.

PW-144The world model does not compose players: the frozen kernel reads only team four-vectors + QB + seennext

Owner: Claude · Lane: unassigned · Priority: 0

Why this matters

T-06 (lineup -> unit function) and T-02/T-03/T-05 player skills are marked done and 44,507 player_latents rows exist, but sim.py:_X / meta.json features of release k51f2563814f06fca7f79 take only team off/def pass/rush latents, qb_epa and seen. Rusher/target/blocker/coverage/kicker vectors never reach the rollout, so injuries and lineup changes below QB cannot move a price, and the backtest measured team-level EWMA ratings, not the designed world model. This is the most likely reason the sim carries no information beyond the closing line (corr(model-close, actual-close) ~ -0.08).

Dependency

2026-09-29 21:00 MT injury capture reaching week 4 (data/cache/injuries_2026.parquet holds weeks 1-3 only as of 2026-09-27); after it fires, read inj_snapshot/avail_snap captured_at for week=4 team rows and re-run role_unit_sensor_coverage.py against week 4 -- if H100 coverage moves off 0/1 for WAS@IND, re-run the step 2-proxy design (or a real as-of design if a genuine pre-kickoff lineup is now capturable) on that data point

Evidence and next step

GATED SEQUENCE (reframed 2026-09-25 02:16 after Codex's challenge: T-06 composes ONLY the QB, and its own natural-experiment validation fails, slopes 2019-22 .526 and 2023-25 .340 vs pooled .808). Do not wire unvalidated units into _X. (1) Per role (rusher, target, blocker, pass-rush, coverage, kicker): a unit map from player_latents with its own natural-experiment test (the player missing vs present, same team, adjacent weeks), predeclared, passing on held-out seasons. (2) As-of expected-active lineups at each lock (needs timestamped injury/depth sensors; see residual_candidates_2026.md). (3) Only validated roles enter _X in a NEW release candidate; refit, E-01 on the new release with as-of lineups, E-03 vs k51f on identical games. Re-examine T-06's done state (QB-only, validation failing). Do not modify k51f.

PW-149Podcast transcription lock can sit stuck for hours while the watchdog reads 'ok'done

Owner: Codex · Lane: — · Priority: 0

Why this matters

2026-09-25: the shared transcription lock was stuck 03:48-11:40 MT with nothing running, blocking all Mac podcast transcription; the podcast lane counts 'lock busy' as a run, so nothing alarmed.

Evidence and next step

Fix: (1) the lock records its holder pid and start time; a holder that is not running, or older than its run budget, is reclaimed automatically. (2) The watchdog counts 'lock busy' as NOT a run once it has lasted longer than one run budget, and alarms. Prove both red: plant a stale lock with a dead pid and confirm reclaim + alarm.

PW-150Sim game totals run 3.5-5 points under real on held-out seasons: locate the points gap drive by drivedone

Owner: Codex · Lane: — · Priority: 0

Why this matters

Found while S-14 was open (commits 2ace20ea..f69eba8d). Totals are the market where the model is furthest off; a 3.5-5 pt bias is larger than the edge anyone needs.

Evidence and next step

Measured so far, held out: total yards within 10 of real but the sim total is 3.5-5.2 points under real; pass share +0.06 over real comes from drift_shift.json inside a state; rush yards per carry 4.1 vs real 4.2-4.5; clock split fixes snaps (162 -> 151 a game) but not totals; run-yardage refit moves rush yards but the total only +0.6. Next: drive-level decomposition of the points gap (points per drive, red-zone TD rate, FG vs TD mix, drives per game) on 2025 held out, then one candidate fix measured paired against the current default. ACCEPTANCE: the gap is located to a named drive-level quantity with a held-out interval, and one candidate fix is measured (adopted only if paired total error improves with CI excluding 0).

PW-152Research findings never reach the served model: 6 adopted hypotheses and 2 L-07 promotions are all outside what Sean is serveddone

Owner: Codex · Lane: — · Priority: 0

Why this matters

Status page 2026-09-25: 69 findings, 6 adopted, 0 live. Adopted hypotheses Q-13, Q-14, Q-19, B13, B41, XD-S07 are not among the served v1.0 model's 78 features; L-07's 2 promotions feed the research simulator, not the served forecasts. So the research arm's measured gains have zero effect on live accuracy.

Evidence and next step

Build the path from an adopted finding to the served model: (1) a candidate served model = v1.0 + the adopted findings as features, fit on <=2025 exactly like v1.0; (2) served-as-shadow every refresh beside v1.0 on the same games, so each finding gets a live A/B; (3) promote the candidate only when its held-out 2024-25 backtest AND its live shadow both beat v1.0 with intervals excluding 0 on the page's metrics; (4) record each promoted change in buildsite/research-served.json automatically with its backtest gain, so the status page's hypothetical-vs-actual table fills itself. ACCEPTANCE: the 6 adopted findings run as a live shadow with per-finding hypothetical and actual gain on the status page; a promotion decision recorded with numbers.

PW-168Workers share one git checkout: give every build run its own git worktreedone

Owner: Claude · Lane: — · Priority: 0

Why this matters

Incident #1 (ops/incidents/INCIDENT-LOG.md): build workers and agents ran git stash on the shared checkout and pulled ~200 files of other sessions' uncommitted work (09-27, 09-28).

Evidence and next step

Each build-loop run works in its own `git worktree` on a short-lived branch and lands with a fast-forward or rebase onto main, then removes the worktree. The shared checkout is used only by interactive sessions. ACCEPTANCE: two concurrent workers editing the same file cannot lose each other's work (planted test); a stash inside a worker affects only that worker's tree; no regression in landing rate over one day.

PW-169A logged-in capture that returns 0 rows when rows are expected must alarmdone

Owner: Codex · Lane: — · Priority: 0

Why this matters

Incident #4: Yahoo current-matchup and Rotowire captures returned 0 rows for 6.5 h on 09-28 with no alarm.

Evidence and next step

For every scheduled capture with an expected-row contract (Yahoo current matchup 32 rows, Rotowire, others found by grep), a run with 0 rows, or below contract, opens a CTRL-02 incident and a watchdog dead row with the reason. A skip because a capture is recent is not a failure. ACCEPTANCE: planted 0-row capture goes red and opens an incident; a normal run closes it.

PW-62The reviewer must not be the builderdone

Owner: Codex · Lane: premise · Priority: 0

Why this matters

The operating model: 'the same model may fill several jobs, but never proposes, implements, and approves the same material claim.' Tonight one process proposed the AAR->topics link, built it and approved it; proposed the stacked arm, built it and graded it; and wrote the mouse landing gate AND the checks that gate runs. The arm-vintage artifact — four contrasts at sign z beyond 26 that were one code difference read twice — was caught only because I happened to look again. Nothing structural would have caught it. Name, for each material claim, who proposed it and who approved it, and make them different. See §2.

PW-68Turn the density-ratio classifier into the loss function (GAIL loop)done

Owner: Codex · Lane: s-observation · Priority: 0

Why this matters

Gemini's re-ranked #1 and its explicit WHAT TO DO FIRST, after Sean pushed it off the coward's consensus.

Evidence and next step

X-03 built a density-ratio classifier that separates real from simulated game states essentially perfectly, and we used it once as an AUDITOR and never again. Gemini: 'Instead of just measuring the disjointness, close the loop: the simulator must minimize the scalar error while maximizing the probability that the classifier flags the simulated trajectory as real.' That is Generative Adversarial Imitation Learning, and it avoids hand-defining a loss for 33 fields over 150 plays. MIDTERM EXAM, its words: wired as the discriminator loss, the classifier's AUC drops from ~1.0 to ~0.5 on simulated 10-step sequences — the kernel has learned football physics rather than a shortcut to the scalar. ITS MECHANISM FOR WHY THIS IS THE ROOT CAUSE, which is worth keeping: grading only the terminal dot product is a credit-assignment failure. 'A model can achieve exactly 15.4 FP by simulating one 80-yard touchdown and 60 consecutive kneels.' Unpenalised intermediate states let the transition kernel learn alien shortcuts, and rolled out autoregressively those errors compound off the manifold. Related: PW-56, PW-57, V-01.

S-16S-16 — Trajectories are content-hashed per (game, lock, release); issuance is frozen with the existing done

Owner: Codex · Lane: unassigned · Priority: 0

Why this matters

Plan item S-16, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Trajectories are content-hashed per (game, lock, release); issuance is frozen with the existing row-lock discipline. | Replays bit-for-bit. FILES THE PLAN NAMES: research/playgrain/sim_store.py MISSING: sim_store.py No prior board item ever mentioned this. CURRENT DELIVERY WORK: REOPENED 2026-09-24 delivery audit: prospective shadow and baseline locks now exist, but sim_store is not their common issuance path. Add shape-aware hashes, verified loading, atomic concurrent creation and actual issuance clocks; wire prospective issuer and pregame archive to one ID. Test corruption, concurrent writes, late horizon claims and exact replay. Preserve every existing issuance; see docs/institute/PROSPECTIVE-SHADOW-2026-09-24.md. REOPENED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): state said done while this detail already said REOPENED, so takeable.py skipped it and no worker could take it. CRITICAL-PATH STEP 2, FIRST: freeze ONE kernel release -- one content hash over the post-c51817b2 model heads (research/playgrain/models/*), the sim.py git sha and the PW_* environment flags, copied into a release directory so the concurrent refits (w5 X-05, w7 S-15) cannot move it, recorded with its id in data/sim_store/releases.json. Every price (N-04) and backtest (E-01, E-03, S-06) cites it. ACCEPTANCE: Sim(models_dir=<release>) replays a seeded game bit-for-bit, releases.json names the id, and the id is written onto E-01. Then the delivery work above.

V-02V-02 — A settled week keeps the number filed before kickoffdone

Owner: Codex · Lane: serving · Priority: 0

Why this matters

Week 1 prints 384.40 under ACTUAL, PLAYERWEEK and YAHOO — the same figure three times, both sides. Week 2 does it on our side and on the opponent's YAHOO column while PLAYERWEEK correctly holds 417.63. Weeks 3 and 4 are right. Storing a forecast with a clock is pointless if the page overwrites it with the result once the week settles. EXAM: week 1 shows three distinct figures and V-01 stops reporting it.

V-04V-04 — The nine stub tabs on the build monitordone

Owner: Codex · Lane: construction · Priority: 0

Why this matters

Eleven sections, and nine are stubs of about 200 characters: Commission & Portfolio, World Model Lab, Football Intelligence, Behavior & Qualitative, Data & Provenance, Markets & Portfolio, Experience Studio, Operations & Learning, Forecast Accuracy Directorate. Only Construction and Maintenance carry anything. Each one either gets the content its title promises or is removed — a navigation item that leads to 200 characters is a lie about what the machine knows. EXAM: no tab under the stub threshold, measured by V-01 extended to the build site.

X-05X-05 — Key numbers: the mass at 3, 6, 7, 10, 14 and 17 on margin, and at 37, 41, 44, 47 and 51 on totaldone

Owner: Codex · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item X-05, lane kernel, drift, decision, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Key numbers: the mass at 3, 6, 7, 10, 14 and 17 on margin, and at 37, 41, 44, 47 and 51 on totals, against history by rules era. | Every ratio in 0.90–1.10. FILES THE PLAN NAMES: research/playgrain/drift/shape_test.py MISSING: shape_test.py No prior board item ever mentioned this. NEXT (CTO + Codex, 2026-09-25 03:33): predeclared same-seed A/B of drift_shift ON vs OFF on the frozen release k51f (a scratch copy of the sidecar; never edit k51f in place). Same slate (2024 and 2025 REG, or a fixed 48-game subset first), same seed, n=300, QB = prior-starter as in the E-01-QB run. Report per arm: pass attempts, snaps, drives per team-game vs observed; mean total error; margin calibration slope; RMSE. PASS for 'drift_shift off (or refit on scramble-clean targets) is better' = total error closer to 0 AND slope closer to 1 on BOTH seasons. The winner becomes a NEW release candidate for E-01/E-03 vs k51f on identical games. OWNERSHIP 04:16 MT: Codex is changing drift/k51f_shift_ab.py to reuse the ON arm's saved E-01-QB draws (after proving bit-identity) so only OFF cells use CPU. Workers: don't edit that file tonight. NEXT QUESTION (Codex + CTO, 05:10 MT 2026-09-25, after the predeclared drift_shift screen FAILED on 2024): the sim scores 1.71-1.80 points per drive vs ~2.11 real while running more drives. Trace which state/transition rates cause it — 3rd-down conversion, red-zone TD vs FG, turnover/punt rates by field position — on REAL game states (the action-probability audit method of afbfc44e), with PW_CLOCK_SPLIT and PW_DRIFT_SHIFT at their defaults. Diagnose first; do not tune on the 48-game A/B. Any fix becomes a named release candidate for a full-origin E-01 vs k51f. A/B RESULT 05:3x MT (k51f_shift_ab.json, complete): predeclared verdict FAIL. 2024: OFF total err +5.56 -> +2.32, slope 1.12 -> 1.42 (SE .6). 2025: OFF total err +2.21 -> -1.29, slope 1.29 -> 1.21 (SE .5), margin RMSE 12.84 -> 13.06. Read: drift_shift OFF roughly fixes totals (slight overshoot in 2025); margin effects are inside noise at 48 games. Workers: take the NEXT QUESTION above (points per drive), diagnosis first. LAPTOP HOURS: heavy sims only 00:00-06:00 daily and 08:00-17:00 Mon-Fri (Mountain). Between 06:00 and 08:00, read-only/light diagnostics only.

C-13C-13 — Fill `published_at` and `effective_at` on the 9,039 receipts, then on every table without a clocdone

Owner: Codex · Lane: sensors · Priority: 1

Why this matters

Plan item C-13, lane sensors, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Fill `published_at` and `effective_at` on the 9,039 receipts, then on every table without a clock (125 of 187), in order of use. FILES THE PLAN NAMES: research/playgrain/sensors/evidence_clocks.py MISSING: evidence_clocks.py No prior board item ever mentioned this.

D-01D-01 — Fair prices from the calibrated distribution for every market on the board; exact push handling done

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item D-01, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Fair prices from the calibrated distribution for every market on the board; exact push handling at integers; edge against the *early* number. FILES THE PLAN NAMES: research/playgrain/decide/price.py MISSING: price.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10. REOPENED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): the pricing math is done (decide/test_price.py passes) but its only measurement scored the broken pre-c51817b2 2008-09 walk-forward ('worse than coin flip') -- INVALID. REMAINING: re-measure edge_vs_early on E-01's re-run (S-16 release) for 2023-2025, the early-line seasons; for weeks 4 and 5 compute edge vs the Monday-Wednesday early number (early_lines, C-04) on N-04's prices -- per game: fair price, early number, edge -- written where D-05's sheet reads it.

D-05D-05 — The sheet: wager, line, walk-away, stake, in window order. Offline, on a phone.done

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item D-05, lane kernel, drift, decision, due Tue 6 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The sheet: wager, line, walk-away, stake, in window order. Offline, on a phone. FILES THE PLAN NAMES: research/playgrain/decide/sheet.py, playerweek/pwa.py MISSING: sheet.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10. REOPENED FOR DELIVERY 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): sheet.py is built; the deliverable is the week-4 sheet on Tue 09-29 and the week-5 sheet on Tue 10-06 from N-04's prices and D-01's edges, stakes $0 while G4 is red (E-03), each sheet stating which gate is red. Depends on N-04 and D-01. GEMINI 1/10 (2026-09-25, staking): write the sizing rule before G4 can open — fractional Kelly (<= 1/4) on the calibrated probability, shrunk by posterior width, capped per bet and per week against the $100k bankroll. Stake stays $0 until G4.

E-08E-08 — Weekly prospective issuance at the TUE lock, frozen beside v1 and the rival, graded the followindone

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item E-08, lane kernel, drift, decision, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Weekly prospective issuance at the TUE lock, frozen beside v1 and the rival, graded the following Tuesday. Two shadow slots, as the schematic specifies. FILES THE PLAN NAMES: research/playgrain/evalx/issue_shadow.py, evalx/grade_week.py MISSING: issue_shadow.py, grade_week.py No prior board item ever mentioned this. REOPENED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): FALSE DONE for the pricing path. It issues only v0.1, POWER and DYN (all teamstate family); the world model is not a shadow slot, and only week 3 was ever issued (forecasts: 3 versions x 16 rows, recorded 09-22 09:04). REMAINING: grade N-04's WORLD version beside the other three; issue v0.1/POWER/DYN for weeks 4 and 5 at the TUE lock; run grade_week for week 3 on Tue 09-29 and week 4 on Tue 10-06 (CRPS, log-loss, ATS vs early per version) into data/results.duckdb.

PW-100Probe an untested limit (cannot) in commit:938e33e9done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:938e33e9: "If pgrep itself cannot be run the row assumes the worse case, because a check that cannot look must not clear an alarm." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-101Probe an untested limit (out-of-scope) in commit:89994e3fdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:89994e3f: "pytest is not installed in .venv so tests/test_board_is_plan.py did not run; out of scope for this change (render.py carries no plan-sync dependency) and gate.py was open both before and after with no plan-of-record edit made." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-102Probe an untested limit (out-of-scope) in commit:6e1730dddone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:6e1730dd: " overstates it (PW-70 has no family), but the substance holds: PW-70 is a different, unowned item on shard 6 (70 % 8), genuinely out of scope for T-02 on shard 2." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-103Probe an untested limit (out-of-scope) in commit:e1766d30done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:e1766d30: "10 of 13 warehouse tables now covered; the 3 refused ones need a versioned/append-only redesign before they can be gated, out of scope for this item's named files." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-106xwalk's season/gsis_id ranking is nondeterministic -- reruns silently swap which roster snapshot winsdone

Owner: Codex · Lane: sensors · Priority: 1

Why this matters

Found while verifying C-13's clock fix on xwalk (playerweek/identity.py), not by probing an assertion.

Evidence and next step

The `ranked` CTE in identity.SQL picks one roster_raw row per (season, gsis_id) with `ROW_NUMBER() OVER (PARTITION BY season, gsis_id ORDER BY <count of non-NULL ids> DESC)` and no further tiebreak. VERIFIED directly, not assumed: ran the UNMODIFIED SQL twice back-to-back against one identical in-memory snapshot of roster_raw/dp_ids -- 147 of 33,898 xwalk rows differed between the two runs, same code, same inputs, zero relation to any change made this session. Whenever two roster snapshots for the same player-season tie on id-fill count, which one xwalk keeps depends on DuckDB's internal ordering, not the data -- so a full rebuild can silently change which pfr_id/yahoo_id/etc a player carries, with no error and no signal. Needs a deterministic final tiebreak column (e.g. the roster row's own capture order, or an explicit sort key) added to the ORDER BY.

PW-107Probe an untested limit (out-of-scope) in commit:53b6dc03done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:53b6dc03: "F-08 stays open -- the binding blocker is the joint correlation matrix, not the QB coverage gap, and closing it needs a copula/joint structure over the six-slot matrix, out of scope for this unit." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-108Probe an untested limit (cannot) in commit:06c4f524done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:06c4f524: "Falls back to the live warehouse if a clean copy can't be taken inside 120s rather than failing the job outright." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-109Probe an untested limit (out-of-scope) in commit:bbabdf70done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:bbabdf70: "Filed as a finding, not fixed: whether its capture_health check is a design decision, out of scope here." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-111Probe an untested limit (cannot) in commit:c970bb94done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:c970bb94: "S-18: a fresh 2025 kernel rollout at HEAD plays totals of 67 against a real 41 -- the live simulator is not a calibrated kernel since the 09-22 S-08 refit (committed evals show it in time order: 42/44 before, 77/88 after), so the kernel-vs-rating bar cannot be re-read yet and no grind is registered." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-114Probe an untested limit (cannot) in commit:668f3540done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:668f3540: "PW-112: the s remainder cannot be settled' claim probed -- half true." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-115Probe an untested limit (cannot) in commit:4f96e73cdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:4f96e73c: "A finding that never proposes anything cannot reach the model, cannot be refereed, and cannot be wrong in a way that costs anything." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-130Probe an untested limit (banned-floor) in commit:c2a23b59done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:c2a23b59: "E-01: re-run origins 2021 (accepted) and 2022 (rejected at noise floor) on k51f2563814f06fca7f79; sweep halted at the CTO's gate, not loosened 2022 misses 1.5pt/0.30 by 0.10pt/0.010; bias SE 0.81 so the bound is ~1.85 SE." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-131Probe an untested limit (out-of-scope) in commit:694f292fdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:694f292f: "Rows that fall out of scope are withdrawn, and the archive queue is exactly the top T1 episodes." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-140Kernel release snapshot is incomplete: policy_target.py missing; rosters.json order loses player alignmentdone

Owner: Codex · Lane: unassigned · Priority: 1

Why this matters

J-11 reproduced the ATL-GB issuance bit for bit only with the live tree's policy_target.py; the release k51f2563814f06fca7f79 does not capture it, and rosters.json is written sorted so a release-only rebuild mis-orders 165 player arrays.

Evidence and next step

Fix kernel_release.py to capture every module sim.py imports (policy_target.py included) and to store roster order as used, cut a new release, and extend tests/test_full_trace.py to rebuild from the release alone with no live-tree reads. Do not modify k51f... in place (it is referenced by sealed issuances); record the gap on it.

PW-141build-loop credits another lane's commit as its own 'progress kept' after an alarm overrundone

Owner: Codex · Lane: unassigned · Priority: 1

Why this matters

01:27-01:29 MT: w1, w4, w5 and w7 each logged 'unit 1 overran the alarm but committed 28dbf4b1 — progress kept', but 28dbf4b1 is the N-12 agent's lead_curve commit; the workers produced nothing. The overrun check looks for ANY new commit, so a stalled unit reads as progress.

Evidence and next step

In bin/build-loop.sh, attribute an overrun only to commits whose message cites an item from THIS worker's brief (or whose author session is this unit's witness lease), and count others as a failed unit. Add a test that a foreign commit during an overrun is not credited.

PW-142com.playerweek.transcribe-heavy calendar job did not fire at 00:00 or 01:00 MT on 09-25done

Owner: Codex · Lane: unassigned · Priority: 1

Why this matters

launchd showed runs=1 (a manual start at 00:03) after the 19:56 plist rewrite; the midnight window's scheduled starts never happened. A schedule is not evidence of completion.

Evidence and next step

Find why the StartCalendarInterval entries did not fire (reload timing, plist shape, launchd state) and add a check that the job ran in each window (log line per firing) that the watchdog reads. Note: a deliberate hold was set 01:07-06:00 MT 09-25 for compute; firings during a hold must log 'held', not silence.

PW-148playprob.fit fits on on_roster=1, the gameday ACT flag -- an outcome-time column; refit on all rows and score through forecast_deliverydone

Owner: Codex · Lane: forecast · Priority: 1

Why this matters

Found while measuring PW-69. A column with no decision-time clock cannot enter a fit (rule 2): on_roster is roster_status='ACT' for THAT week (build.py:250), the gameday active list published ~90 minutes before kickoff, so playprob.fit (playprob.py:114) trains on players who were not gameday-inactive.

Evidence and next step

Signature: rostered Questionable rows no-play 0.4-1.9% in 2020-25 but 15-29% in 2016-18 and 9.4% pooled; 2,500 NONE rows with on_roster=0 played 0.2%; 798 of the 950 Out/Doubtful rows since 2020 are INA. Refitting the SAME cells on all rows (arm inc_all) beat the shipped fit -0.0044 [-0.0050,-0.0039] pinball on snap share, 5/6 seasons (Questionable -0.0033 [-0.0048,-0.0018] 4/6). WHAT TO DO: (a) refit playprob.fit without the on_roster=1 filter, walk-forward; (b) score through forecast_delivery on served rows in POINTS, not snap share -- the snap-share gain is not yet a points gain; (c) establish whether served rows are gameday-filtered at decision time, which decides how much of this reaches a shipped number; (d) check the other exams whose universe is on_roster=1 for the same conditioning (backtest.py:26 UNIVERSE, arms.py, aar.py). Evidence: docs/build-evidence/PW-69.md.

PW-153Per-section freshness stamps (ff116765) are not rendered on any served pagedone

Owner: Codex · Lane: — · Priority: 1

Why this matters

Sean, 2026-09-22: "I can't tell if the build and the research organization sections are getting updated." ff116765 added per-section stamps in buildsite/board.py (_asof), but the served pages are built by buildsite/desks.py, which never calls it. Measured 2026-09-25 21:58 MT: 0 section-asof elements in 15 dist pages while section data ages span 1.0 to 28.9 min under one header stamp.

Evidence and next step

Port _asof() (mtime of the data a section renders; warn >15m, bad >45m; live=True for render-time sections) into buildsite/desks.py for the machine map (buildsite/state.json) and any section reading data/site-findings.json, or route those sections through board.py. Read bin/site_markers.py first. Verify against the LIVE url after deploy, and plant a stale file to see it go red (rule 1). Found by PW-77.

PW-154Probe an untested limit (cannot) in commit:79e12960done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:79e12960: "H-215 released OPEN unrun: defmodel's defender lines are issued nowhere, so that layer cannot be graded (BLOCKED row added)." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-155test_zone_joins.py reports rc=1 on a fully passing run -- its own summary line crashesdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

Found while triaging PW-92's 168 baseline-not-green files: tests/test_zone_joins.py's 4 real checks (kickoff-zone columns measure their declared zone, nothing in the warehouse is unzoned, no SQL joins two zones naively, the live read-lag guard) all pass -- the FAIL list is empty when the crash happens. The crash is in the success-path summary string at line 215: sorted({r['measured'] for r in measured}) raises TypeError('<' not supported between NoneType and str) because that set mixes None and strings, so the test always exits rc=1 whether or not its checks pass, and J-08's mutation harness cannot exercise it (rc=1 at baseline).

Evidence and next step

tests/test_zone_joins.py:215, inside the f-string building `note` after FAIL is confirmed empty: sorted({r['measured'] for r in measured}) over a set containing both None and str values. Fix: sort with a key that handles None (e.g. sorted(..., key=lambda x: (x is None, x))) or drop None before sorting. Reproduced live 2026-09-27 against the unmodified file.

PW-158test_roster_positions.py: MINE constant no longer matches league.rosters() keys, so the live standings page can never pick the user's own teamdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

Found while triaging PW-92's 168 baseline-not-green files, timeout slice: test_roster_positions.py reproduces live (unmodified file) as 'week 1 lineup: 0/0 slots filled, 0.0 points'. Root cause: playerweek/league.py:31 MINE = "Slop King Dynasty" (old team display name), but league.rosters() now keys by manager first name (Sean, Landon, Rick, Clint, Brian, Matt, James, Remington, Doug S, Bryan) -- "Slop King Dynasty" is not among them.

Evidence and next step

season.py:499-501 season.project_team(con, MINE, ...) returns an empty TeamWeek whenever rosters().get(manager) misses, silently rather than erroring. The same stale constant also breaks Standing.is_mine at season.py:720,738 (m==MINE is never true), which playerweek/pwa.py:1645/2621/2626 reads to pick plan_pts vs projected_points for 'my' row on the live standings page -- currently always falls through to the non-mine branch. Fix: update MINE to the current manager key (or resolve it dynamically instead of hardcoding a display name that renames).

PW-160Probe an untested limit (not-possible) in commit:882de132done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of measuring it: the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:882de132: "The +0.19 pt oracle ceiling on QB carries is not reachable through post-hoc calibration." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-161Probe an untested limit (out-of-scope) in commit:58efda68done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:58efda68: "Filing the remaining ~39 stale-measurement files as individual PW- items stays out of scope per Sean explicit deferral." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-162Probe an untested limit (out-of-scope) in commit:5dfbf278done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:5dfbf278: "PW-93: extend captured_at UTC audit to drain.py, confirm landed-JSONL round trip is safe Grepped every playerweek/ module for captured_at + INSERT (the scope two units ago called out of scope for one unit): aar.py, bhypo.py, forecast_delivery.py, statusboard.py only read or comment on the column." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-163Probe an untested limit (out-of-scope) in commit:c93859fddone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:c93859fd: "schedule.install() already contains the fix (unload any job.hour>=0 plist while CALENDAR_OWNED is False) but a full install() run unloads all ~55 calendar descriptors at once, out of scope for one unit on a live Sunday slate." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-164Probe an untested limit (out-of-scope) in commit:6dc9c6abdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:6dc9c6ab: "Full finding and cross-checks in docs/evidence/PW-157-2026-09-27-dispatch-line-ground-truth.json, including why this does NOT mean CALENDAR_OWNED should flip back to True (44/46 clock jobs are still legitimately since the outage) and an unrelated, pre-existing test_dispatch.py failure found live (7" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-165Probe an untested limit (out-of-scope) in commit:c6b0bc2fdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:c6b0bc2f: "New finding: the dispatch-line-ground-truth evidence file (18:42 MT) found 7 jobs double-registered with both launchd and the poller (arm-forecast-refresh, arm-live-refresh, fpleague-refresh, o3-followup, q01-daily-haiku, served-scorecard, suite-check) and called it out of scope." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-167Probe an untested limit (out-of-scope) in commit:3ac92243done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:3ac92243: "Read-only check against the live data/nimo-status.json confirms the new row already reads dead against the real, unresolved last_contact clock (~23h old at check time) -- Nimos active claim, out of scope here." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-172Probe an untested limit (out-of-scope) in commit:9cc76fb0done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:9cc76fb0: "tests/test_cadence.py fails on this tree independent of this change (pre-existing calendar_as_list list/dict bug, confirmed via git show HEAD:playerweek/schedule.py) -- out of scope." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-174E-03 early-line clock: load_early refuses pre-previous-week snapshots -- re-run canonical G4 and the other consumers on the fixed loaderdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

H-310 (referee UPHELD dbc9b45b): odds_history 2023-25 is one capture a day, so the 'nearest T-100h' early line predated the previous week's last kickoff on 146 of 831 games, and any model that has seen week w-1 earned free CLV there. G4 condition 2 and every early-line grade (calibrate.py, edge_early.py, N-11 edge_screen, pw143_replay) read that loader.

Evidence and next step

DONE IN THIS ITEM'S FIRST COMMIT (market lane, 2026-09-29): market_diag.load_early(clean_clock=True) refuses any 60-200h snapshot captured at or before (previous REG week's last kickoff + 4h) and picks the nearest-T-100h among the rest; clean_clock=False reproduces the old loader on all 831 pinned-parquet games. FIXTURE tests/test_early_clock_guard.py: RED on the unfixed loader (market_diag sha b663b9c4 at f0228a21: 3 failures + 1 error of 5 -- a Sunday wk2 game kept its pre-MNF snapshot, a TNF wk2 game took the pre-MNF -2.5 instead of the clean -1.5), GREEN after (5/5); tests/test_release_market_diag.py still 5/5. MEASURED (experiments/2026-09-29T2219Z-market-g4-clean-clock.json, prereg 3f79f953): G4 condition 2 on clean clock, k51f 2024-25: spread CLV -0.040 [-0.147,+0.070] n 503; total +0.090 [-0.004,+0.190] n 504 -- FAILS (canonical legacy: -0.087 n 543 / +0.130 n 544, replayed exactly). 35 games re-picked, 40 dropped. G4 RED on all three conditions either way. A prior-3-season league-mean total earns +0.077 of the +0.090 total CLV; the model's partial slope given it is +0.001 [-0.054,+0.052]. REMAINING (owner E-03): (1) DONE 17ca54bd: release_market_diag.py --update-canonical, g4_verdict.json now cites n 503/504. (2) re-run the consumers that read load_early -- DONE cc79e20b for decide/calibrate.py, decide/edge_early.py, evalx/edge_screen.py (N-11): population shrank ~7-8%, no conclusion changed (docs/evidence/PW-174-2026-09-29-consumer-rerun.md); evalx/pw143_replay.py's actual replay left to the grind registered in 19787faa (pw143_replay_2025.json, grind_done), not yet complete. (3) DECIDED 2026-09-29 (build-w6): no change to market_diag.g4_verdict()'s condition-2 pass rule -- it does not move today's RED verdict (spread already fails outright at -0.040 regardless of any total-side rival adjustment), it is not what the 2026-10-10 stake reads (that gate is chain-report.py's prospective 2026 CLV, not this historical grade), and g4_verdict() is plan sec.6's own pure function, so redefining condition 2 is a plan-level call left for its own review. The rival control (+0.077 of +0.090 total CLV is league-mean reversion; partial slope +0.001 [-0.054,+0.052]) stays disclosed in the frozen experiment for that review. Full reasoning: docs/evidence/PW-174-2026-09-29-item3-rival-decision.md. ACCEPTANCE: g4_verdict.json condition 2 cites n 503/504 (or the re-run's n), and the fixture stays green -- MET. CLOSED 2026-09-29 (build-w6): pw143_replay's grind (19787faa) completed 2026-09-30T00:00:11Z, grind_done=1 in pw143_replay_2025.json (commit 958a3033) -- population shrank 272->251 games, same ~7-8% shrinkage as the other three consumers, PW-143's own verdict unchanged (rejected, non-verifiable-historical-benchmark). All three named sub-items of the remaining work are done; nothing left to grind or measure. tests/test_early_clock_guard.py + tests/test_release_market_diag.py 10/10 green. See docs/evidence/PW-174-2026-09-30-pw143-replay-clean-clock.md.

PW-175B-01 remainder: coach_dim 2026 head coach is last year's coach for ARI/ATL/BUF -- extend coordinators to 2026 and rebuild; tests/test_coach_dim_2026_hc.py red todaydone

Owner: Codex · Lane: — · Priority: 1

Why this matters

H-194 construction refines (2026-09-28T2320Z, 2026-09-29T2318Z): coordinators.build stops at last=2025, so coach_dim covers 2026 from games_raw only (544/544 rows source=games_raw_fallback), and nflverse games.csv still names the fired coach at week 1 for ARI (Gannon, world LaFleur), ATL (Morris, world Stefanski), BUF (McDermott, world Brady), and spells LV 'Kubliak'. The 2026 new-head-coach class reads 7 teams, the world's is 10. Every 2026 reader of coach_dim (coach features, hurdle_pred's coach category, policy fits keyed to head_coach) inherits it; the H-194 09-30 read protects itself with a per-read override and nothing else does.

Evidence and next step

RED TODAY: tests/test_coach_dim_2026_hc.py fails 3/3 on the live warehouse (named-team week-1 coach; exact new-HC set = ARI ATL BAL BUF CLE LV MIA NYG PIT TEN; zero 2026 games_raw_fallback rows). GREEN DEMONSTRATED ON A SCRATCH COPY (research/scientist/run/construction_h194_coachdim_fixture.py): playerweek.coordinators.collect(first=2026,last=2026), unmodified, returns 32/32 infobox head coaches matching the ESPN witness; appended to coordinators and passed through the unmodified coach_dim.build, the test goes 3/3 green, all 544 2026 rows become coordinators_csv, 13,934 pre-2026 rows are identical, and exactly 68 team-weeks change (ARI/ATL/BUF/LV x 17). A rebuild with unchanged inputs stays red and is row-identical to live. THE BUILD: coordinators to 2026 (fetch 2026 only; keep the 2016-2025 CSV rows byte-for-byte), save_csv, rebuild coach_dim, add the test to the suite. OPEN JOINT: the infobox is a live page with no per-edit receipt clock -- stamp the fetch time and do not relabel weeks already kicked off when an in-season firing is written in. Record: research/scientist/experiments/2026-09-29T2318Z-construction-h194-coachdim-fixture.json

PW-176Coordinator infobox parser merges Eagles 2026 OC and DC fieldsdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

The 2026 Eagles infobox puts adjacent template parameters on one line. coordinators._field currently captures the next key as part of off_coach, contaminating the coordinator record.

Evidence and next step

Planted source |off_coach=Sean Mannion|def_coach=Vic Fangio}} currently yields OC Sean Mannion|def_coach=Vic Fangio}} and DC None. Bound the field at the next template parameter, preserve valid wiki links, and verify the Eagles 2026 record and historical coordinator rows. Source: commit 30b1b1bd.

PW-177Measure current promotion backlog evidencedone

Owner: Codex · Lane: — · Priority: 1

Evidence and next step

Consume PW-74 exact existing grinder/progress receipt and current backlog trend history. Measure whether production-to-promotion backlog changed and record the bounded next source action; do not start repeated model polling.

PW-180Verify week 4 injury snapshot eligibility for player compositiondone

Owner: Codex · Lane: — · Priority: 1

Evidence and next step

Read actual week 4 inj_snapshot/avail_snap captured_at and run role_unit_sensor_coverage.py on week 4. Record exact coverage and pre-kickoff eligibility; preserve absent historical clocks and scientific negatives.

PW-181Land the submitted isolated worker delivery for PW-168done

Owner: Codex · Lane: — · Priority: 1

Evidence and next step

Inspect exact data/build-land-retries receipt for PW-168 and original worker branch; use guarded ff-only landing if current source permits. Preserve dirty work and report exact blocked paths.

PW-182Repair forecast publisher team opponent validationdone

Owner: Codex · Lane: — · Priority: 1

Evidence and next step

Consume exact failed run 20260930T200425-ff198957 state.json and frozen product/site-products.json: team 6 week 4 opponent 2. Inspect /Users/seano/dev/playerweek-frontend/playerweek/site_export.py apply_team_products rejecting team product 0 wrong opponent. Repair source mapping in the owning checkout, preserve frozen product and opponent correctness, run focused publisher validation, and submit actual source commit. Do not rerun capped INC-99cbafe132 attempts or bypass the product gate.

PW-183Verify current suite source repair admissiondone

Owner: Codex · Lane: — · Priority: 1

Evidence and next step

Compare current terminal suite report and data/suite-repair-route-ready.json against exact current source hash. Use existing guarded suite preflight to verify actual failing test roots without spending model repair budget. Repair source-bound route evidence only when deterministic verification warrants it; preserve negative tests, original repair caps, owner holds and product gate.

PW-186Picks for every unplayed game this seasondone

Owner: Sean · Lane: — · Priority: 1

Why this matters

Sean, 2026-10-01: decisions are written only one week ahead, so the season has no forward slate of picks to look at or stake against.

Evidence and next step

Write a pick (side and total, with the model margin, total and their sd) for every unplayed 2026 game, not only the coming week. The frozen v0.1 model already has margins for all 272 games in game_bets (read-only check 2026-10-01: 272 rows, 272 distinct game_id; columns pred_margin, pred_total, pred_margin_sd, pred_total_sd). Today the decision writer stops one week ahead. ACCEPTANCE: every 2026 game with no actual_margin has exactly one current pick on the served product, each traceable to its game_bets row and model; weeks beyond the next are labelled as early picks; a planted test fails if any unplayed game lacks a pick.

PW-187efficiency.fit changes when player_week is merely reordered: make efficiency.py:182 order-free (H8 red, H-274)done

Owner: Claude · Lane: construction · Priority: 1

Why this matters

Invariant H8 has been red since 2026-09-25 and every R&D lane that shadows player_week stops on it (H-336: 19 of 42 records on 09-30/10-01 were stop records). Re-measured 2026-10-03 (invariants.py run started 08:32Z) at 19764e2de: reordering identical player_week rows changes 10,685 of 10,940 efficiency.fit(season <= 2022) player-rates. The repair is recorded and measured (H-274) but nothing carried it to a builder: grep H-274 was 0 on this board.

Evidence and next step

efficiency.py:182 pools each player's totals under ANY_VALUE(position), which picks whichever row the scan meets first; that label moves his position's prior posr. Fix: an order-free rule, recorded as arg_max(position, season * 100 + week) in research/scientist/experiments/2026-09-27T2323Z-construction-h274-repair-check.json (H8 red->0 on <=2022 and <=2025; walk-forward 2019-25 paired MAE +0.00004 pts, CI spans 0: a nil model change); mirror at defmodel.py:114 (32 -> 0 order-moved defenders through 2026). Latest vs modal position is a football choice; order-freedom is the defect. RED FIXTURE: tests/test_efficiency_position_order_free.py -- synthetic, no warehouse, one two-position player in two physical orders; at HEAD it FAILS (84 of 130 fitted rates move) and its control (the shipped module with the recorded edit applied in memory) PASSES. ACCEPTANCE: that test goes green without editing its assertion or fixture; research/scientist/invariants.py H8 reads [ok]; check_H8 widened to season <= 2025 and a defmodel <= 2026 reorder check added, each seen red at the old line; the next walk-forward (wf_std.py) is regenerated at the repaired HEAD and its paired MAE vs the prior preds is reported (expected within +-0.001).

PW-195Probe an untested limit (cannot) in commit:f9e75c479done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:f9e75c479: "A full record now carries a ninth body section, , between Decision and Rerun (FULL_SECTIONS); a preregistered record does not, since the bridge can't be reported before the result exists." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-196Probe an untested limit (cannot) in commit:18f943edcdone

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:18f943edc: "product-truth: world-tue waives week 5 (Tue seal missed, cannot be issued late) and requires only the current week (INC-39b44f2002) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> PW-Run: r20261008T095535-build-w3-ae0058 PW-Employee: build-w3" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-205Decide: the text box is the answer, not a side note beside Approve/Rejectnext

Owner: Claude · Lane: — · Priority: 1

Why this matters

Sean, 2026-10-15: 'I didn't like that it was approve or deny even though the box was text.' On a yes/no card the text box is labelled 'Optional note', so a real answer ('yes, but do X') has nowhere to go except a note the bridge may ignore.

Dependency

Work is edits plus a deploy in ~/dev/decide, a separate repo outside this worker worktree; needs a decide-repo seat to take it. Not started.

Evidence and next step

In ~/dev/decide src/page.js and src/index.js: make typed text the primary answer on every card. If he types, the button becomes Send and the text is the answer (bin/decide.py pull must act on it, including conditional yes/no in words). Approve/Reject remain one-tap shortcuts only when the box is empty. Verify on the live phone page and that pull hands the text to the asking seat. Same design is specified for the life app in ~/dev/life-manager/docs/APPROVALS.md.

PW-207E-03/G4: land the qualified 2021-22 Wednesday (T1d) rung as a second early-line source in market_diag.load_earlydone

Owner: Codex · Lane: — · Priority: 1

Why this matters

G4 condition 1 (flat -110 return at EARLY prices, interval excluding 0) is graded on odds_history 2023-26 only, and every early-line question since 2026-10-09 has ended at a power stop. The T1d rung (data/land/odds-historical/<season>-w<week>-T1d.jsonl, receipt sha256 883fcd4d...) adds 538 clean 2021-22 games (1,354 for 2021-25), qualified in research/scientist/experiments/2026-10-10T0730Z-market-t1d-rung-g4-power.json: same-book within-3h identity with odds_history 92.6% of 6,785 pairs. Register row H-535 (market lane).

Evidence and next step

SPEC: in load_early add the T1d rung for seasons 2021-22 only (no overlap with odds_history), keep G4's 60-200 h lead window, apply the four frozen checks -- market_diag.info_cutoffs clean clock; pre-kickoff; capture identity |captured_at - asked_for| <= 1 h; kickoff identity |commence_time - games_raw kickoff| <= 1 h -- median across books, home-favoured sign (spread_home negated, as for odds_history). Refuse every 2026 rung file (14 of 46 matched 2026 rows fail capture identity: files fetched days before the asked clock hold another week's snapshot). The reference loader is research/scientist/run/market_h534_sbr_reversal.t1d_moves (asserts n=1354 for 2021-25); reuse its logic, do not fork it silently. NEGATIVE FIXTURES AS TESTS (extend tests/test_early_clock_guard.py; each must be proven RED by disabling its check once, and the disabling recorded in the commit): (1) shift the rung's week label +1: the 831 rows that join on (season, week, home, away) all have lead >= 165.8 h, so 0 rows may enter with lead < 120 h; (2) same-book identity: T1d vs odds_history within 3 h agree on the spread point in >= 90% of pairs; (3) a planted 2026 T1d file is refused; (4) a planted snapshot captured before the previous week's last kickoff + INFO_LAG_H is refused (existing clean-clock rule must also hold on the new source); (5) a planted commence_time 2 h off games_raw kickoff is refused. ACCEPTANCE: the tests are green and each was seen red; then a G4 condition-1 re-grade on 2021-25 for money_backtest fc_v01 at early prices, reported with n and a 95% interval beside the existing 2023-25 number (expected: a tighter interval, not a green G4; break-even needs about a 1.44-pt edge). Acceptance is the graded interval, never code merged. The market lane does not re-grade G4 itself (coaching sci-market-primary v11).

PW-25Probe an untested limit (cannot) in board:D-04done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from board:D-04: "**Opened this Tuesday with paper entries.** It cannot be backfilled, and it costs nothing." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work. ID RE-PREFIXED 2026-09-22. This item was filed as C-25, which collides with the plan's own C-01..C-18 namespace. A lane reading 'C-19' could not tell a residual-ledger proposal from a plan item, and the two mean entirely different things. Locally-generated items carry PW- now; the plan owns the bare letter prefixes.

PW-26Probe an untested limit (banned-floor) in board:J-10done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from board:J-10: "Delete the stale "irreducible" text; fail the build on `irreducible\|noise floor\|ceiling` applied to football." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work. ID RE-PREFIXED 2026-09-22. This item was filed as C-26, which collides with the plan's own C-01..C-18 namespace. A lane reading 'C-19' could not tell a residual-ledger proposal from a plan item, and the two mean entirely different things. Locally-generated items carry PW- now; the plan owns the bare letter prefixes.

PW-28Probe an untested limit (out-of-scope) in commit:31c2db51done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:31c2db51: "migration is out of scope for this item, and there's no .github/workflows to" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work. ID RE-PREFIXED 2026-09-22. This item was filed as C-28, which collides with the plan's own C-01..C-18 namespace. A lane reading 'C-19' could not tell a residual-ledger proposal from a plan item, and the two mean entirely different things. Locally-generated items carry PW- now; the plan owns the bare letter prefixes.

PW-29Probe an untested limit (cannot) in commit:0611cf72done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:0611cf72: "snapshot exists for a moment and cannot be backfilled; settlement grades" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work. ID RE-PREFIXED 2026-09-22. This item was filed as C-29, which collides with the plan's own C-01..C-18 namespace. A lane reading 'C-19' could not tell a residual-ledger proposal from a plan item, and the two mean entirely different things. Locally-generated items carry PW- now; the plan owns the bare letter prefixes.

PW-31Lead: separate two worlds for his snap sharedone

Owner: Codex · Lane: — · Priority: 1

Why this matters

The residual ledger named his snap share as a known unknown worth 0.3332. This lead proposes an observation that would tell two competing explanations apart. It is a LEAD, not a design: it passed the airlock (schema + at least one resolving citation) and nothing more.

Evidence and next step

{ "two_worlds": [ "The player is an every-down foundational starter whose usage is immune to offensive game script.", "The player is a situational package specialist dependent on down, distance, and game flow." ], "separating_observation": "The percentage of first-quarter, first-and-ten plays the player was on the field for.", "where_to_get_it": "The free 'nflverse' public participation dataset (pbp_participation), specifically by parsing the 'offense_players' column for the player's GSIS ID.", "citations": [ "https://nflreadr.nflverse.com/reference/load_participation.html", "https://github.com/nflverse/nflverse-data/releases/tag/pbp_participation" ], "confidence": "high", "_dead_citations": [] } ID RE-PREFIXED 2026-09-22. This item was filed as C-31, which collides with the plan's own C-01..C-18 namespace. A lane reading 'C-19' could not tell a residual-ledger proposal from a plan item, and the two mean entirely different things. Locally-generated items carry PW- now; the plan owns the bare letter prefixes.

PW-36AAR: conversion error is 32% of our total, and it is forecast's layerdone

Owner: Codex · Lane: forecast · Priority: 1

Why this matters

The after-action review attributes 32% of all forecast error we have actually made to conversion — 2,804 rows, mean |error| 11.43. That layer belongs to the forecast lane: efficiency: turning opportunity into points. This is not a counterfactual about what would be worth knowing; it is the error we made, decomposed.

Evidence and next step

Worst positions by count: QB (1,271), WR (693), RB (516). READ THE DECOMPOSITION BEFORE PROPOSING ANYTHING. aar_attribution carries e_variance, e_regime, e_volume, e_share and e_conversion per forecast, so the question is not 'how do we get better' but 'why does THIS component dominate for THESE positions'. AND CHECK IT AGAINST THE RESIDUAL LEDGER, which answers a different question — what a variable would be worth if known perfectly, rather than what we actually got wrong. Where the two disagree is the interesting case: an error we keep making on a variable the ledger says is cheap means the cost is in our method, not in the unknown. A variance-labelled error is the honest residual of a stochastic process and may be correct; the question there is whether our INTERVALS cover it, which is calibration's layer, not whether the point estimate was wrong.

PW-37AAR: share error is 8% of our total, and it is allocation's layerdone

Owner: Codex · Lane: allocation · Priority: 1

Why this matters

The after-action review attributes 8% of all forecast error we have actually made to share — 1,032 rows, mean |error| 8.14. That layer belongs to the allocation lane: who gets the ball, inside the room. This is not a counterfactual about what would be worth knowing; it is the error we made, decomposed.

Evidence and next step

Worst positions by count: WR (491), RB (286), TE (237). READ THE DECOMPOSITION BEFORE PROPOSING ANYTHING. aar_attribution carries e_variance, e_regime, e_volume, e_share and e_conversion per forecast, so the question is not 'how do we get better' but 'why does THIS component dominate for THESE positions'. AND CHECK IT AGAINST THE RESIDUAL LEDGER, which answers a different question — what a variable would be worth if known perfectly, rather than what we actually got wrong. Where the two disagree is the interesting case: an error we keep making on a variable the ledger says is cheap means the cost is in our method, not in the unknown. A variance-labelled error is the honest residual of a stochastic process and may be correct; the question there is whether our INTERVALS cover it, which is calibration's layer, not whether the point estimate was wrong.

PW-38Probe an untested limit (out-of-scope) in commit:4a9ba165done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:4a9ba165: "Left alone -- out of scope for" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-39Probe an untested limit (not-worth) in commit:a2e6a739done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:a2e6a739: "Still not worth building." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-40Probe an untested limit (cannot) in commit:16e1274bdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:16e1274b: "cannot be recovered later." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-41Probe an untested limit (cannot) in commit:60a5ac67done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:60a5ac67: "Week 3 would otherwise go unfiled and the ledger cannot be backfilled." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-42Probe an untested limit (banned-floor) in commit:837568d5done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:837568d5: "irreducible noise" (contradicting this project's own D-345 correction a few" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-43Probe an untested limit (banned-floor) in commit:837568d5done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:837568d5: "irreducible|noise floor|hard ceiling|inherent limit (reusing bin/premise-" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-44Probe an untested limit (cost-estimate) in commit:9895266bdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:9895266b: "fetching the 372 columnsa multi-month" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-45Probe an untested limit (banned-floor) in commit:e4521890done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:e4521890: "("a measurement-procedure noise floor: the spread you get from refitting the" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-46Probe an untested limit (out-of-scope) in commit:412242d1done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:412242d1: "Wiring the mechanism into drives.py is out of scope for this item -- it" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-47Probe an untested limit (out-of-scope) in commit:153a9251done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:153a9251: "One adjacent gap recorded in the result, not fixed here (out of scope for this" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-52Give the QB a passing leg in the attribution chain -- conversion is not one bucketdone

Owner: Codex · Lane: forecast · Priority: 1

Why this matters

PW-36 found that attrib.decompose() (playerweek/attrib.py) never substitutes truth into a QB's passing production -- pass_attempts, pass_yards, pass_tds, interceptions ride unchanged from the forecast through every intermediate score to conversion = m2 - m3. That makes conversion the ONLY label a QB miss outside its interval can get, whatever actually moved: a pick-six and a missed deep shot are attributed identically. QB is 45% of all conversion-labelled rows (1,271 of 2,804) partly because of this, not because per-attempt efficiency is where QB forecasting is worst -- and nobody can tell which it is until the chain can tell them apart.

Evidence and next step

trace.py's CHAIN (playerweek/trace.py:26-34) lists nine fields, all receiving or rushing. Add a passing leg the same shape as the existing two: hold the QB's TRUE pass_attempts, substitute true per-attempt rates (completions/att, yards/att, tds/att, interceptions/att) the way _rates()/_line() already do for targets and carries, and add a fifth decompose() step between volume and conversion so the QB's own passing splits into a volume term (pass attempts) and a genuine per-attempt efficiency term. THIS CHANGES STORED aar_attribution ROWS. Any change to decompose() must be followed by attrib.classify() and must still pass tests/test_aar_attrib_fresh.py's per-row exact-reproduction check (TOL=1e-9) and the identity check in decompose() itself (the components must still sum to total_error to 1e-6). Re-derive the conversion/regime/volume/share shares after the change -- QB's share of 'conversion' should fall once passing has somewhere else to go, and that shift is the thing to report, not just that the code compiles.

PW-53Probe an untested limit (banned-floor) in commit:b21879f1done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:b21879f1: "PW-42: probe irreducible-noise claim in commit 837568d5 -- survives, its (J-10) message, which quotes the stale" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-54Probe an untested limit (cost-estimate) in commit:0f5799bbdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:0f5799bb: "PW-44: probe the multi-month/fetch claim on P-06 -- refuted premise-scan's claim was already false on the day it was made: P-06" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-58Stack the arms S->B->Q instead of pooling themdone

Owner: Codex · Lane: forecast · Priority: 1

Why this matters

Sean, 2026-09-22: "SBQ should be stacked S-->B-->Q because B is kinda a version of statistical data, whereas Q is actually thinkings like judgement and sentiment."

Evidence and next step

Today an arm is a FEATURE SUBSET handed to one ridge, so SBQ is an independent fit that can land anywhere relative to S — and does: B+Q move 99.8% of forecasts, hurting 6,944 and helping 5,377 (sign z=+14.1) for a net +0.0072 MAE whose interval crosses zero. Most of that movement is the whole fit shifting because ten columns entered the penalty, not signal. A stack stores delta_B and delta_Q per player-week, so 'we cut him on the language and were wrong' becomes a row you can query; a row with no Q signal gets exactly zero instead of a fit-wide perturbation. research/stacked_arm.py is built and unmeasured.

PW-59Wire the podcast corpus into featuresdone

Owner: Codex · Lane: q-language · Priority: 1

Why this matters

12.3M transcribed words reach nothing.

Evidence and next step

playerweek/transcripts.py was built for the newsletter brief and filters to rostered players; no transcript-derived column exists in share.ALL_FEATURES. features.py declares three and all are built=False: stated_role_change ('a reporter or coach saying the staff wants him more involved — LEADING rather than lagging'), coach_intent, insider_confidence. The Q arm is five columns of injury-report boilerplate while this sits unread, so every verdict on Q so far is a verdict on the boilerplate. Needs the 2024-25 backlog transcribed first — com.playerweek.transcribe is now on a 55-minute clock.

PW-60Probe an untested limit (cannot) in commit:11cc5974done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:11cc5974: "six open gaps named so they cannot be mistaken for oversights." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-67Rotate proposer and critic so one provider's style is not our priordone

Owner: Codex · Lane: fable · Priority: 1

Why this matters

'Gemini 3 Pro finds cross-domain mechanisms; a second model converts them into state, equations, falsifiers and gates.' Tonight Gemini proposed AND judged, reached what it called the coward's consensus twice, and needed Sean to push it back both times. Step 3 of our pipeline asks it to judge itself. A different model should hold that chair. See §7.

PW-69Sever B and Q from the Ridge; apply them as bounds on the transition matrixdone

Owner: Codex · Lane: forecast · Priority: 1

Why this matters

Gemini's re-ranked #2, replacing its own residual-stacking proposal after the research found no published evidence for it.

Evidence and next step

Its diagnosis of the Ridge survives and is sharp: 'Because B and Q are sparse, to make them move the terminal forecast when they do fire, the optimizer must assign them massive weights. To pay for those under the L2 budget, it is mathematically forced to shrink your highly predictive S-weights. You are actively taxing your dense physics engine to pay for sparse linguistic adjustments.' And the geometry error: 'an injury does not evenly subtract 3.5 points; it means a 30% chance of leaving the game early and 70% of playing normally. A flat regression cannot output a bimodal distribution; it can only smear it into an inaccurate mean.' THE FIX IS SEAN'S OWN SYNTHESIS, which Gemini calls the only correct frame: language and decisions are evidence about PARAMETERS of the simulated distribution, not a shift on a mean. An injury designation caps a max snap-transition probability; it does not subtract points. ITS MIDTERM EXAM IS STALE AND MUST BE RESTATED. Gemini wrote: 'the +14 z-score sign split vanishes completely.' That z=+14 was OUR number and it was an artifact — four arms compared across two code vintages. Re-measured from one vintage on 15,022 paired player-weeks, every contrast's interval crosses zero and every sign split is near even (B without Q -0.0066 [-0.0194, +0.0061]). There is no +14 to make vanish. The honest exam: B and Q applied as simulation bounds beat the S-only baseline by an interval that EXCLUDES zero — which they have never done in any arrangement yet. Supersedes the pooled-vs-stacked framing in PW-58.

PW-70The covariance compiler: price a correlated prop off the joint rolloutdone

Owner: Codex · Lane: decision · Priority: 1

Why this matters

Gemini's re-ranked #3, and the thing it says is our only structural monopoly.

Evidence and next step

'By reading all markets from one joint distribution, you natively price the cascading tail covariances that sportsbooks' static copulas miss.' This is S-15 with an exam attached. MIDTERM EXAM, its words: the compiler prices a correlated derivative prop — trailing QB pass attempts OVER plus leading RB rush yards OVER — directly from the joint distribution of rollouts, yielding a price that DIVERGES from the sportsbook's implied parlay odds while MATCHING the historical play-by-play baseline. Both halves are required: diverging without matching history is just being wrong. Its own pre-mortem on this, kept because it is the thing that would kill it: the verdict dies if a sportsbook's closed-form static copulas already capture >95% of cross-market tail risk, in which case the world model is computationally inferior and structurally redundant.

PW-77Probe an untested limit (cannot) in commit:ff116765done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:ff116765: "One stamp cannot be true for all of them." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-78Probe an untested limit (cannot) in commit:605cea4ddone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:605cea4d: "A findings file that cannot be read counts as -1, not" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-79Probe an untested limit (not-possible) in commit:13a9dc8edone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:13a9dc8e: "Now it is hard, in both the picker and the brief: while ANY plan item is open in a shard, a PW- item is not available to take." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-80Probe an untested limit (cannot) in commit:605cea4ddone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:605cea4d: "A findings file that cannot be read counts as -1, not 0: an audit that never ran is not a clean site, and defaulting to review on a read error is how the site job would quietly stop." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-81Probe an untested limit (out-of-scope) in commit:153a9251done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:153a9251: "One adjacent gap recorded in the result, not fixed here (out of scope for this item): nothing in bin/ or launchd actually invokes playerweek.suite on a schedule, so is true only when someone runs the suite by hand." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-82Probe an untested limit (out-of-scope) in commit:412242d1done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:412242d1: "Wiring the mechanism into drives.py is out of scope for this item -- it needs both teams simulated jointly per takeaway, which the result text already flags as S-14s own status of " PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-83Probe an untested limit (out-of-scope) in commit:575058efdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:575058ef: "Not wired into grade_week.pys simpler z-test in this unit -- out of scope for the scoring library itself." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-84Probe an untested limit (out-of-scope) in commit:93bd4e60done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:93bd4e60: "These are raw source-ingest tables needing a per-table ETL change each, out of scope for one unit; the script ranks them by row count for the next pass." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-85Probe an untested limit (out-of-scope) in commit:4a9ba165done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:4a9ba165: "Left alone -- out of scope for this item, worth a careful look next time someone touches rivalgap.py." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-86Probe an untested limit (out-of-scope) in commit:31c2db51done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:31c2db51: "Residual: no other script calls results.record() yet -- that repo-wide migration is out of scope for this item, and theret on GitHub Actions." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-87Probe an untested limit (cannot) in commit:0611cf72done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:0611cf72: "The pre-kickoff availability snapshot exists for a moment and cannot be backfilled; settlement grades games that already finished and loses nothing by waiting." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-88Probe an untested limit (cannot) in commit:16e1274bdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:16e1274b: "That is the only property that cannot be recovered later." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-89Probe an untested limit (cannot) in commit:9800600ddone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:9800600d: "And C-56 established the constraint that shapes the whole plan: nflverse publishes participation POST-SEASON (2024 landed 2025-09-04), so 2026 cannot be fetched and must be imputed from snap counts, depth charts, personnel tendencies and pbp role ids, trained where truth is known." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-90evidence_clocks.py's clock probe treats ANY TIMESTAMP column as a clock, over-counting coveragedone

Owner: Codex · Lane: sensors · Priority: 1

Why this matters

Found while working C-13 (shard 5, 2026-09-22): CLOCK_PAT in research/playgrain/sensors/evidence_clocks.py matches column NAMES against words like published/ingested/modified, but duckdb_clockless_tables() also does `or "TIMESTAMP" in ty` -- so any TIMESTAMP-typed column counts as a clock regardless of what it means. `base` reads as clocked solely because its `kickoff` column (game time, not an ingestion/capture clock) is TIMESTAMP-typed; it has no ingested_at and none of build.py's base_sql() columns carry one.

Evidence and next step

Rule from CLAUDE.md: 'check that a column means what its name says' (games_raw.total was the historical example). This probe violates that rule against itself -- it is the C-13 measurement instrument, and its 95/201 'have a clock' count is inflated by every table that happens to carry an unrelated TIMESTAMP column (kickoff, birth_date-as-timestamp if any, capture windows, etc). The true count of tables genuinely missing an ingestion/publish/capture clock is HIGHER than the 105/201 currently reported, not lower. FIX: drop the `or "TIMESTAMP" in ty` clause entirely and rely on CLOCK_PAT matching the column NAME only (it already covers ingested/captured/published/effective/modified/etc plus a bare `_at` suffix, which is the actual convention this project uses for every clock column added under C-13). Re-run and diff the missing list before and after to see how many tables were false-negative-missing (i.e. wrongly marked 'have'). Does not change C-13's current unit of work (lagged had no candidate column of any kind), but it does mean the 105-of-200 fraction should not be trusted as a stopping condition or progress measure until this is fixed.

PW-91pw-now.py's STALE_MIN check mass-expires every live worker at midnightdone

Owner: Codex · Lane: operations · Priority: 1

Why this matters

Found while clearing the marker after X-04 (shard 4, 2026-09-23): `bin/pw-now.py --clear` reported "cleared build-w4" with no `expired` list moments after marking live, then a later `--clear` in the same unit wiped ALL six other workers' entries (w2/w3/w5/w6/w7, timestamps 23:33-23:47 the prior day) at once, leaving now.json empty.

Evidence and next step

pw-now.py parses each entry's `at` as bare HH:MM and reattaches TODAY's year/month/day before computing age against STALE_MIN=35 minutes. The date rolled from 2026-09-22 to 2026-09-23 at local midnight between two workers' writes; every entry stamped 23:xx on the old day, reparsed against the new day's date, computed as ~23 hours old (or, since the code only accepts -5<=age<=35, simply outside range) and got dropped as stale in one shot -- not because the writers died, but because midnight happened. This is the exact failure the STALE_MIN block's own comment warns against ("a marker that outlives its writer... puts a reader on a hunt for a lane that does not exist") inverted: here live lanes vanished from the board while genuinely still running. FIX: compare against a stamped epoch/ISO timestamp instead of bare HH:MM, or carry the date alongside the time so age is computed across a day boundary correctly. Self-healing in practice (every worker overwrites its own entry on its next unit), so no data was permanently lost, but the machine map read empty for a stretch of real activity right at 00:00 local -- worth fixing before the next midnight, not urgent tonight.

PW-94A concurrent git reset on the shared working tree wipes every lane's uncommitted work at oncedone

Owner: Codex · Lane: operations · Priority: 1

Why this matters

Measured 2026-09-23 ~02:25 MT: git reflog on this shared checkout shows two `git reset` events landing back at HEAD (E-07's commit, 8fa32e8b) within the same minute. `git show --stat` on that commit confirms it touched only 3 files (construction.json, now.json, mde_table.py) -- it did NOT itself carry the files that vanished. Something else ran a reset against the tree while other lanes' edits sat unstaged in it, and every one of those edits was gone afterward: this unit's own in-progress edit to playerweek/regime.py/build.py and its construction.json write (grep for the added text came back empty after the reset, present again only after this unit redid the edit from scratch); and, separately, everything this session's git-status snapshot showed modified at session start (bin/chain-report.py, six buildsite/dist/*.html pages, buildsite/chain.json, research/scientist/HYPOTHESES.md, research/scientist/run/h150_room_domain.log) is gone from `git status` now with no commit that explains where it went.

Evidence and next step

8 shards share ONE working tree (not 8 worktrees), so any script in any lane that runs a blind `git reset`/`git stash`/`git checkout --`/`git add -A` over the whole tree can destroy another lane's not-yet-committed work with no error and no trace -- the destroyed lane's own unit just silently has to redo the work (costly but survivable) or, worse, records a result for an edit that was never actually committed (silent data loss, not survivable, and matches this repo's own rule 3: verify the object under test is the object that shipped). This is the same class of finding as the git-add-sweeps-staged-index and shared-register-partial-staging lessons already on record, but at the working-tree layer instead of the index layer: no amount of care about staging scope protects an unstaged edit from a concurrent hard reset elsewhere in the fleet. NOT INVESTIGATED HERE (out of scope for the unit that found it): which script actually issued the reset -- every buildloop/autoloop/chain-run entrypoint that runs `git` unconditionally is a candidate and would need to be greped and audited to name the one at fault, and the reflog alone does not carry a process name.

PW-95Probe an untested limit (cannot) in board:PW-93done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from board:PW-93: "bin/weather-capture.py called duckdb.connect() directly, skipping the retry, the honest LockTimeout and D-106's 'a writer yields to the capture' — now routed through warehouse.connect and added to the capture exemption beside availability and availsnap, because a forecast at T-8h cannot be re-fetche" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-96Probe an untested limit (out-of-scope) in board:PW-94done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from board:PW-94: "NOT INVESTIGATED HERE (out of scope for the unit that found it): which script actually issued the reset -- every buildloop/autoloop/chain-run entrypoint that runs `git` unconditionally is a candidate and would need to be greped and audited to name the one at fault, and the reflog alone does not carr" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-97Probe an untested limit (cannot) in commit:129b56f4done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:129b56f4: "Committing them here via an isolated GIT_INDEX_FILE + compare-and-swap ref update so this commit cannot be swept by another shard's concurrent git add -A the way the last two attempts were." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-98Probe an untested limit (out-of-scope) in commit:65eb92cddone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:65eb92cd: "out of scope for the scoring library itself") -- it was true." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-99Probe an untested limit (out-of-scope) in commit:9f460d2ddone

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:9f460d2d: "covariance_compiler and evalx-issue-shadow are other families' dead lanes, out of scope this unit." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-HL3Headline metrics on every page: moneyline, O/U, ten best five-leg parlays, both modelsnext

Owner: — · Lane: — · Priority: 1

Evidence and next step

CLAUDE.md 'The three headline metrics' (Sean 2026-10-10). Produce validated, record-of-origin output (bin/lab-snapshot.py and the main-site contract) with win % and ROI at real prices, n and 95% intervals, held-out and live, for Stats and World; render as the first thing on lab home, main site home and the AARs; verify live at 390px with the Access token.

S-11S-11 — Kickoffs (with the 2024 rule era), punts, returns, blocks, and field goals by kicker rating × didone

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item S-11, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Kickoffs (with the 2024 rule era), punts, returns, blocks, and field goals by kicker rating × distance × wind × altitude × roof. Also extra points, two-pointers and onside kicks. | FG make curve calibrated; no possession starts at a constant 30. FILES THE PLAN NAMES: research/playgrain/kernel/special_teams.py MISSING: special_teams.py No prior board item ever mentioned this. UNBLOCKED 2026-09-25 07:40 MT (Claude): The kicker rating and onside trigger are buildable here, not outside dependencies: fit kicker make-probability by distance from warehouse FG attempts (shrunk to league), and draw the onside ATTEMPT from the empirical rate by score margin x clock in play-by-play. Build both inside this item.

S-15S-15 — The compiler: margin, total, moneyline, team totals, alt lines, any prop, fantasy points under adone

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item S-15, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The compiler: margin, total, moneyline, team totals, alt lines, any prop, fantasy points under any scoring (through `statline.score`, the single scoring function), K and DST. All read from the same trajectory parquet. | No number is published that is not a query. FILES THE PLAN NAMES: research/playgrain/queries.py MISSING: queries.py No prior board item ever mentioned this. CURRENT DELIVERY WORK: REOPENED 2026-09-24 delivery audit: compiler exists but production still serves separate legacy models. Build a backend adapter over ONE sealed trajectory artifact that stores game/player/prop products with the same issuance hash; reject partial/regulation artifacts as full-game products. No frontend calculations or promotion without calibration. See docs/institute/PROSPECTIVE-SHADOW-2026-09-24.md. NARROWED AND RAISED TO 0 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): GAME PRODUCTS FIRST -- margin, total and moneyline (alt spreads at -3/-7 included) read from one sealed sim_store artifact (S-16 release) through research/playgrain/serving_adapter.py, which N-04 calls at issuance. queries.py already reproduces price.py on 2,136 games; the game-market serving path is the remaining work. Player, prop and fantasy products come after and are not part of this priority 0. ACCEPTANCE: one call with (game_id, lock, release) returns fair spread/total/ML prices whose input is the sealed trajectory hash. GEMINI 1/10 (2026-09-25, MC error budget): report the Monte Carlo standard error of every compiled quantity and choose n per market so MC error < 0.1 pt on spread/total and < 0.5% on probabilities; tail markets (alt lines, props) get more n. UNBLOCKED 2026-09-25 07:40 MT (Claude): S-06 is done (measured FAIL, calibration is an improvement, not a prerequisite). Close this item: add the Monte Carlo standard error of every compiled quantity (Gemini clause) to the compiler output, verify on the sealed ATL/GB report, mark done.

T-02T-02 — Regularised adjusted plus-minus at the play grain: a sparse design matrix (players on field × rodone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 1

Why this matters

Plan item T-02, lane ratings, lineups, policies, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Regularised adjusted plus-minus at the play grain: a sparse design matrix (players on field × role) in `scipy.sparse`, solved by LSQR with ridge, targets EPA, success, yards, pressure, completion. Hierarchical prior by position, as-of each week, exponential decay. Output `player_latents(gsis_id, role, valid_time, mu, sd, n_eff)`. | Adds play-level log-loss to the kernel, held out, above the parsimony tax. FILES THE PLAN NAMES: research/playgrain/talent/rapm_fit.py MISSING: rapm_fit.py No prior board item ever mentioned this.

V-03V-03 — Every week renders, or declares itself empty by designdone

Owner: Codex · Lane: serving · Priority: 1

Why this matters

Weeks 15-18 render 529 characters against a 3,492 median — 15% of a normal week. A fantasy playoff opponent that is not yet determined is a REAL unknown and saying so is correct behaviour; silently rendering a stub is not, and neither is silencing the check. The declaration must be in a form the audit can read. EXAM: no week below 40% of the median without a declared reason.

V-05V-05 — One declared contract for what a page may showdone

Owner: Codex · Lane: serving · Priority: 1

Why this matters

playerweek/gameserving.py already governs the model's published columns: a field is declared, its producer named, and an absent column emits NULL rather than changing the shape. The pages have no such contract, so a column can silently change what it means — which is exactly what V-02 is. EXAM: adding a rendered field forces a decision, the way adding a serving column already does.

X-02X-02 — Add `form_off`, `form_def` (leave-one-play-out EPA in this game) to `LATENT`; fit the distributidone

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item X-02, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Add `form_off`, `form_def` (leave-one-play-out EPA in this game) to `LATENT`; fit the distribution of form given ratings; the simulator draws it once per trajectory per unit. | Simulated margin sd 14.2 ± 0.3; yards per play within 0.05 of real. FILES THE PLAN NAMES: research/playgrain/drift/game_effect.py, features.py MISSING: game_effect.py No prior board item ever mentioned this.

X-04X-04 — Simulation-based calibration of a small vector of logit offsets against drive and game moments fdone

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item X-04, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Simulation-based calibration of a small vector of logit offsets against drive and game moments from `drive_week`, under common random numbers. The moments are plays per drive, third-down rate, TD/FG/punt rates, totals, margin sd and key-number masses. | All moments within tolerance at once. No more cancelling errors. FILES THE PLAN NAMES: research/playgrain/drift/sbi_calibrate.py MISSING: sbi_calibrate.py No prior board item ever mentioned this.

A-02A-02 — P(active) → P(plays \| active) → snap share \| plays. LightGBM plus isotonic, fitted at each locdone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 2

Why this matters

Plan item A-02, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

P(active) → P(plays \| active) → snap share \| plays. LightGBM plus isotonic, fitted at each lock (TUE, WED, THU, FRI, SAT, T-90); the feature list follows this table. | Log loss beats carry-forward *and* the pre-kickoff consensus (the OR-12.2 rows). FILES THE PLAN NAMES: research/playgrain/lineup/avail_hurdle.py MISSING: avail_hurdle.py No prior board item ever mentioned this.

A-05A-05 — For each trajectory, draw who is active, the role depth chart, and the on-field set for each perdone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 2

Why this matters

Plan item A-05, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

For each trajectory, draw who is active, the role depth chart, and the on-field set for each personnel grouping. Returns the `Lineup` the kernel consumes. | Simulated snap shares against realised: MAE by position, published. FILES THE PLAN NAMES: research/playgrain/lineup/lineup_sampler.py MISSING: lineup_sampler.py No prior board item ever mentioned this.

B-03B-03 — Personnel grouping (11/12/21…) given state and roster. This selects the on-field set from A-05.done

Owner: Codex · Lane: ratings, lineups, policies · Priority: 2

Why this matters

Plan item B-03, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Personnel grouping (11/12/21…) given state and roster. This selects the on-field set from A-05. FILES THE PLAN NAMES: research/playgrain/policy/policy_personnel.py MISSING: policy_personnel.py No prior board item ever mentioned this.

C-10C-10 — Coach and coordinator press conferences within 24 h. Podcast priority queue over the 461,382 inddone

Owner: Codex · Lane: sensors · Priority: 2

Why this matters

Plan item C-10, lane sensors, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Coach and coordinator press conferences within 24 h. Podcast priority queue over the 461,382 indexed episodes: team-beat shows first, 72 h before each game, `whisper-large-v3-turbo` at ~27× realtime. Proper nouns are resolved against the 450,604-row roster table ("George Corloftis" → Karlaftis). FILES THE PLAN NAMES: research/playgrain/sensors/pressers_capture.py, sensors/podcast_transcribe.py MISSING: pressers_capture.py PRIOR EVIDENCE on the old board: C-10

C-14C-14 — The injury-report ladder and depth charts "cannot be walked backwards" only through one endpointdone

Owner: Codex · Lane: sensors · Priority: 2

Why this matters

Plan item C-14, lane sensors, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The injury-report ladder and depth charts "cannot be walked backwards" only through one endpoint. nflverse git history and the Wayback Machine do hold them. Rebuild Wed/Thu/Fri practice status for 2016–2025. FILES THE PLAN NAMES: research/playgrain/sensors/wayback_backfill.py MISSING: wayback_backfill.py PRIOR EVIDENCE on the old board: C-14

F-05F-05 — Add the built-and-locked but unused PRE_KICKOFF blocks: home, rest, division, travel_km, tz_shifdone

Owner: Codex · Lane: truth and the fantasy path · Priority: 2

Why this matters

Plan item F-05, lane truth and the fantasy path, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Add the built-and-locked but unused PRE_KICKOFF blocks: home, rest, division, travel_km, tz_shift, elevation, roof, surface, OL continuity, pressure allowed/made, six turnover-cause rates. | One feature family per run, recorded FILES THE PLAN NAMES: none all present No prior board item ever mentioned this.

F-09F-09 — The shipped DST ranking is *exactly* the market's. Rebuild it from unit ratings. | DST rank corrdone

Owner: Codex · Lane: truth and the fantasy path · Priority: 2

Why this matters

Plan item F-09, lane truth and the fantasy path, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The shipped DST ranking is *exactly* the market's. Rebuild it from unit ratings. | DST rank correlation to market < 1 FILES THE PLAN NAMES: kdefmodel.py all present No prior board item ever mentioned this.

PW-147Kernel candidate from T-09: lighter novice-QB prior, and a shrunk kicker term (tau=30)done

Owner: Codex · Lane: unassigned · Priority: 2

Why this matters

T-09 held-out proxies. For n<4 QBs, K=0 beats K=4 on next-game EPA in both 2024 and 2025. Kicker tau=30 beats the pooled FG curve in 2025 (-0.0036 log loss [-0.0058,-0.0014]); in 2024 the CI crosses 0.

Evidence and next step

Both are unit proxies on already-used seasons. Predeclare the setting and test it on a matched frozen-kernel replay, then prospectively on 2026 weeks 4+. It needs a new release id. ACCEPTANCE: game-margin and FG-points effect measured with CIs against k51f on identical games.

PW-194Transcript trait-claim extraction pipeline (stages 1-3 of TRANSCRIPT-MINING-PROGRAM)next

Owner: Claude · Lane: — · Priority: 2

Why this matters

Sean, 2026-10-06: transcripts are 'a gold mine we haven't tapped yet' for B and Q features in both models. 57.7M+ transcribed words reach zero features today. This is data processing, so it never pauses (CLAUDE.md THE FOCUS).

Evidence and next step

Build stages 1-3 of docs/research/TRANSCRIPT-MINING-PROGRAM.md: (1) trait-claim extraction into a typed contract output with subject, trait, polarity, intensity, conditions, speaker and role, source, quote span and said-at timestamp, using a recall filter (local model on the Desk GPU allowed, see ~/compute/README.md) then a strong model on candidates; (2) trait canonicalisation into a taxonomy; (3) point-in-time per-subject aggregation with source reliability and independence weighting. Run it over the full corpus. Acceptance: - a 200-passage hand-checked gold set reaching precision >= 0.85 and recall >= 0.7 on trait claims - every claim traceable to its quote span - an as-of query returns only claims said before the cutoff Validation and value tests (stages 4-6) belong to the B/Q and testing research lanes.

PW-197H-450 — Trace legal downs, possession and score beneficiaries in the sealed World kernel.done

Owner: — · Lane: — · Priority: 2

Why this matters

Tested-positive research card H-450 (research/scientist/HYPOTHESES.md, status CLAIMED) has no implementation unit on the board yet. [card node=D.drives track=B rollup_wpp=0.15 p_real=0.7 cost=3 residual=world-boundary-unverified] **Trace legal downs, possession and score beneficiaries in the sealed World kernel.** MECHANISM: a turnover or fourth-down stop changes who owns the ball and the next drive; one wrong orientation corrupts both player credit and later points. DESIGN: candidate-bound D02-D04 and S02-S03 fixtures with mirrored home/away traces, retained/lost fumbles, fourth down, try returns, and ledger score conservation; plant each wrong transition before accepting the checker. Compare real drives using the same bo

Evidence and next step

Filed by bin/feed-predictive-queue.py from HYPOTHESES.md CLAIMED card. Acceptance is a GRADED RESULT on the shared point-in-time harness (winner % vs Vegas with n and interval, or the relevant metric-tree node) -- never code merged.

PW-198H-454 — Measure the unimplemented World fidelity nodes against real games before claiming a gain.done

Owner: — · Lane: — · Priority: 2

Why this matters

Tested-positive research card H-454 (research/scientist/HYPOTHESES.md, status CLAIMED) has no implementation unit on the board yet. [card node=P.epa track=B rollup_wpp=0.35 p_real=0.55 cost=4 residual=world-lower-node-distance-missing] **Measure the unimplemented World fidelity nodes against real games before claiming a gain.** MECHANISM: play success/explosives, drive boundaries and points per drive generate team points, margin and player volume; margin/total distance alone cannot validate them. DESIGN: candidate/base paired games, discovery <=2023, replication 2024 and heldout 2025; extract real and simulated P.sack/turnovers, P.epa/explosive, D.drives/start, D.ppd/turnovers, T.points, V.availability/role with declared d

Evidence and next step

Filed by bin/feed-predictive-queue.py from HYPOTHESES.md CLAIMED card. Acceptance is a GRADED RESULT on the shared point-in-time harness (winner % vs Vegas with n and interval, or the relevant metric-tree node) -- never code merged.

PW-200H-464 — Does compressing player forecasts into fantasy points discard useful game-score information?done

Owner: — · Lane: — · Priority: 2

Why this matters

Tested-positive research card H-464 (research/scientist/HYPOTHESES.md, status CLAIMED) has no implementation unit on the board yet. [card node=T.points track=A rollup_wpp=0.15 p_real=0.25 cost=3 residual=player-game-bridge-discards-stat-composition] **Does compressing player forecasts into fantasy points discard useful game-score information?** MECHANISM: identical fantasy totals can encode different passing/rushing touchdowns, yards and turnovers; H-391 currently sums proj_points and fits a scalar score map. Preserve the SAME frozen player stat lines and compare a fixed six-stat residual bridge with that scalar blend and a TD-only rival, both against exact v0.1. This is information sufficiency, not H-134's REFUTED scorer-

Evidence and next step

Filed by bin/feed-predictive-queue.py from HYPOTHESES.md CLAIMED card. Acceptance is a GRADED RESULT on the shared point-in-time harness (winner % vs Vegas with n and interval, or the relevant metric-tree node) -- never code merged.

PW-202H-515 — Is World team separation limited by justified uncertainty, miscalibration, or missing score information?done

Owner: — · Lane: — · Priority: 2

Why this matters

Tested-positive research card H-515 (research/scientist/HYPOTHESES.md, status CLAIMED) has no implementation unit on the board yet. [card node=T.points track=B rollup_wpp=0.15 p_real=0.3 cost=2 residual=team-strength-dispersion-versus-information] **Is World team separation limited by justified uncertainty, miscalibration, or missing score information?** MECHANISM: persistent offense/defense quality affects play efficiency, but noisy prior games justify shrinking its conditional mean. H-467 reports R_head=.659 and slope1.135; increasing dispersion alone can worsen an already calibrated forecast. Parent H-467 owns the latent candidates/refit; H-408 is VOID on its drive-count control. Novelty: add calibrated-partial-knowledg

Evidence and next step

Filed by bin/feed-predictive-queue.py from HYPOTHESES.md CLAIMED card. Acceptance is a GRADED RESULT on the shared point-in-time harness (winner % vs Vegas with n and interval, or the relevant metric-tree node) -- never code merged.

PW-203World kernel — next fidelity item from kernel-next-REPORT.mddone

Owner: — · Lane: — · Priority: 2

Why this matters

World kernel fidelity work named in research/strategy/world-2026-10-09/kernel-next-REPORT.md has no implementation unit on the board yet; read that file for the concrete next action before starting.

Evidence and next step

Filed by bin/feed-predictive-queue.py from research/strategy/world-2026-10-09/kernel-next-REPORT.md. Acceptance is a GRADED RESULT on the shared point-in-time harness (winner % vs Vegas with n and interval, or the relevant metric-tree node) -- never code merged.

PW-75CUT: the unattended Deep Think / Deep Research loopnext

Owner: Sean · Lane: operations · Priority: 2

Why this matters

Sean, 2026-09-22: "if I can't trust you to run it in the web then you should cut it from the plan."

Dependency

Sean: nothing — this is a record of a decision, not a request

Evidence and next step

bin/deep-cycle.py drove Gemini's Deep Think and Deep Research through an isolated Chrome profile. Deep Research NEVER WORKED: C-168 records that the profile cannot carry a Google session without a one-time interactive sign-in, and run_research() has always written a FAILED row rather than a success it never saw — which was honest and still meant the feature did not exist. com.seano.playerweek.deep-think is unloaded. WHAT IS NOT CUT: the three-step pipeline Sean runs by hand — prompt 1 Deep Think, prompt 2 Deep Research, prompt 3 the judge, in docs/gemini/. That is a human-run workflow with a real result already (it produced the audit, the four contrasts, and the coward's-consensus correction). The machine automating it unattended is the part that was fiction. TO REVIVE IT the bar is a single observed round trip, not another attempt: one real Deep Research response captured from the browser path. Until then the brief builder in deep-cycle.py stays as dead code with this note on it.

T-05T-05 — Weekly state-space update per player-skill: process noise by age and position, a shock term on rdone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 2

Why this matters

Plan item T-05, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Weekly state-space update per player-skill: process noise by age and position, a shock term on return from injury. The posterior **mean and variance** are stored, and the simulator samples from the posterior, so rating uncertainty becomes game variance. | Calibrated: standardised one-step errors ~N(0,1). FILES THE PLAN NAMES: research/playgrain/talent/kalman_latents.py MISSING: kalman_latents.py No prior board item ever mentioned this.

X-06X-06 — The column-by-column real-against-simulated feature diff that found the 3.2-sd `games_seen` shifdone

Owner: Codex · Lane: kernel, drift, decision · Priority: 2

Why this matters

Plan item X-06, lane kernel, drift, decision, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The column-by-column real-against-simulated feature diff that found the 3.2-sd `games_seen` shift, as a standing gate. | No feature beyond 0.25 sd. FILES THE PLAN NAMES: research/playgrain/drift/feature_shift_diff.py MISSING: feature_shift_diff.py No prior board item ever mentioned this.

A-06A-06 — Role regime changes as a changepoint posterior on snap share (`regime.py` and `kalman_share` exidone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 3

Why this matters

Plan item A-06, lane ratings, lineups, policies, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Role regime changes as a changepoint posterior on snap share (`regime.py` and `kalman_share` exist), not a raw delta. FantasyPros' entire edge sits on players whose situation is changing (+4.753 MSE at 15+ pts of snap movement). | Beats `RECENCY_FEATURES`, which was measured as noise. FILES THE PLAN NAMES: research/playgrain/lineup/depth_roles.py MISSING: depth_roles.py No prior board item ever mentioned this.

B-05B-05 — Box count, rushers, man/zone, shell, given state, offensive personnel and DC. These are the W-00done

Owner: Codex · Lane: ratings, lineups, policies · Priority: 3

Why this matters

Plan item B-05, lane ratings, lineups, policies, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Box count, rushers, man/zone, shell, given state, offensive personnel and DC. These are the W-001…W-014 columns that are "held on disk and read by no feature". FILES THE PLAN NAMES: research/playgrain/policy/policy_defense.py MISSING: policy_defense.py No prior board item ever mentioned this.

B-06B-06 — Who gets the ball. Run: carrier among on-field backs and QB (designed or scramble). Pass: a multdone

Owner: Codex · Lane: ratings, lineups, policies · Priority: 3

Why this matters

Plan item B-06, lane ratings, lineups, policies, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Who gets the ball. Run: carrier among on-field backs and QB (designed or scramble). Pass: a multinomial over the routes on the field, given target-earning ratings, the coverage draw, the depth-of-target distribution and QB tendencies. Identity-free, so it works for any roster. It attacks the −0.5036 usage-share oracle directly. FILES THE PLAN NAMES: research/playgrain/policy/policy_target.py MISSING: policy_target.py PRIOR EVIDENCE on the old board: C-79

C-17C-17 — College play-by-play and player seasons from collegefootballdata.com (free key) for rookie priordone

Owner: Codex · Lane: sensors · Priority: 3

Why this matters

Plan item C-17, lane sensors, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

College play-by-play and player seasons from collegefootballdata.com (free key) for rookie priors. FILES THE PLAN NAMES: research/playgrain/sensors/cfbd_ingest.py MISSING: cfbd_ingest.py PRIOR EVIDENCE on the old board: C-17

PW-113Two lanes are named for a vendor instead of a jobdone

Owner: Codex · Lane: operations · Priority: 3

Why this matters

Sean, 2026-09-23: "What does the label 'Codex' mean? That is a tool name not the job it is doing."

Evidence and next step

Every other lane is named for what it does — s-observation, calibration, decision, serving. Two are named for the CLI that happens to serve them: bin/codex-site.sh and bin/codex-review.sh, surfacing as 'codex' on every marker, log line and chart. THE NAMES ARE experience AND audit. The first owns the product surface; the second independently checks measured claims. Neither is about OpenAI — bin/pick-model.py already fails the audit job over to Gemini or Claude when the outside seat is empty, and the only thing that travels with it is the constraint that a reviewer may never be the builder. A vendor name on a lane makes that failover look like a different lane, which is exactly how five days of Codex silence went unnoticed: nobody was looking for a job called 'audit'. Rename the scripts, the PW_LANE values, the watchdog entry and the markers. Keep the vendor in the MODEL field, which is where it belongs.

PW-34The 15-minute model-fired progress cron is the expensive kind of loopdone

Owner: Claude · Lane: — · Priority: 3

Why this matters

Audit finding, and Sean should know the cost. The update itself reads precomputed files and is cheap in work; it is not cheap in tokens, because each firing lands in this session's full context. It has fired roughly 150 times since it was set.

Dependency

the remaining acceptance clause (the actual token-costing 15-minute cron confirmed stopped/widened with a before/after number, or Sean's explicit decline) is a CTO-seat/Sean action outside this repo and outside the build lane's charter -- searched exhaustively for it in-repo already (see result) and it is not there. Re-check only once Sean or the CTO seat has acted; there is nothing new to search for in the meantime.

Evidence and next step

Everything the update reports is already computed at zero model tokens by launchd -- gate.json, chain.json, watchdog.json, the plan strip on the page. The model adds two lines of judgement per tick, and on a QUIET tick adds nothing. Two honest options: widen to 60 minutes and lose almost nothing, or have the page carry a 'last 15 minutes' delta block (progress.py already computes it) so the update is a thing he reads rather than a thing that is generated. Sean's call on spend, which is why this is filed rather than done.

PW-64The nine stub tabs ARE the nine desks — V-04 is a staffing gapdone

Owner: Codex · Lane: construction · Priority: 3

Why this matters

Commission & Portfolio, World Model Lab, Football Intelligence, Behavior & Qualitative, Data & Provenance, Markets & Portfolio, Experience Studio, Operations & Learning, Forecast Accuracy Directorate — an exact match to D01-D08 plus the accuracy desk, 37 chartered jobs between them. The institution was chartered and never staffed, and nine 200-character pages are the visible shape of that. Work V-04 with this in mind: each tab shows what its desk knows, or the desk is struck. See §4.

PW-65A job is a charter, not a promptdone

Owner: Codex · Lane: construction · Priority: 3

Why this matters

'A versioned charter with inputs, permitted actions, required evidence, timing, and a SUCCESSOR.' Our lanes have briefs — no successor, no independent reviewer, no output contract, no version. LANES.md plus a launchd interval is about half a charter, and the missing half is the accountability half. See §5.

PW-72Record why the NOT FOUNDs are the interesting part, not the discouraging partdone

Owner: Codex · Lane: premise · Priority: 3

Why this matters

The reasoning Gemini reached only after Sean pushed back, worth keeping as doctrine.

Evidence and next step

Three research targets came back NOT FOUND and its first instinct was 'an operational suicide mission'. Its second: 'Public academia is only measuring 1-step context or static divergence. THIS ABSENCE IS THE ALPHA. If academia hasn't mapped the decay of 150-play rollouts, it means public consensus believes it is computationally intractable. The proprietary syndicates who have solved it do not publish.' And on Target C: 'the literature is completely blind to the structural asymmetry you are building.' That inversion is already written into docs/GEMINI-NO-COWARDICE.md and the prompts. This item is to check it holds when we act on it: an unmapped frontier is a candidate moat, and it is also where a solo operator can burn a season. The discipline is that each of these gets a MIDTERM exam short enough to fail fast.

PW-74The register is the logjam: production scaled 4x tonight, promotion did notdone

Owner: Claude · Lane: decision · Priority: 3

Why this matters

Sean, 2026-09-22: "the top half is green, but everything is going to logjam at the register level?" Measured the moment he asked, and he is right.

Evidence and next step

236 hypotheses, 88 OPEN. 819 crumbs closed. 72 findings. 31 proposed changes, 13 still waiting. ZERO claims ever independently reviewed. Tonight the mice went 3 to 10 and the research lanes 4 to 12 — a 4x increase in production with no increase downstream of the register. THE AUTHORITY RULE NAMES THE FIX AND WE HAVE HALF OF IT: 'Promotion comes from INSTRUMENTS and an independent reviewer.' The reviewer is Codex, four slots a day, and site work takes priority on three of them. That cannot clear a queue this size and it should not have to. The instruments are the half that scales: a proposal carrying its own measurement and passing a gate should promote without a reviewer, and the reviewer should spend its scarce slots on the ones instruments cannot settle. ALSO MEASURED: 390 mouse runs found nothing or collided on the claim lock, because crumbs.py --next re-enumerates every surface per mouse behind one lock and took 2m18s with THREE. With ten it is the second bottleneck. Enumerate once into a queue and let mice pop from it.

Q-01Q-01 — Claim schema v2 adds **hedge**, **time anchor**, speaker role and copy ancestry. Q-22 is registedone

Owner: Codex · Lane: sensors · Priority: 3

Why this matters

Plan item Q-01, lane sensors, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Claim schema v2 adds **hedge**, **time anchor**, speaker role and copy ancestry. Q-22 is registered and untested because the fields do not exist. Re-extract all 12,139 news bodies plus pressers and transcripts with the local reader, keeping the verbatim-evidence check. About 46 h of Mac time, overnight. FILES THE PLAN NAMES: playerweek/claims_v2.py MISSING: claims_v2.py No prior board item ever mentioned this. PARKED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): agent units stop (~150 per 72h); the local-reader grind keeps running through bin/grind.py. PASSDOWN §2 measured language at zero value for points. UNPARKED 2026-09-25 07:40 MT (Claude): parked with no recorded reason.

Q-02Q-02 — Settle every claim against what happened. Only 7 are settled, so language's value is *unmeasureddone

Owner: Codex · Lane: sensors · Priority: 3

Why this matters

Plan item Q-02, lane sensors, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Settle every claim against what happened. Only 7 are settled, so language's value is *unmeasured*, not null. Per-speaker and per-outlet reliability as house effects (the XD-B04 house-effect test passed; `consensus.py` shows +4.51 pts/team/season at 4.23σ). Feeds A-02 and B-02. FILES THE PLAN NAMES: playerweek/claim_settle.py, playerweek/speaker_reliability.py MISSING: claim_settle.py, speaker_reliability.py No prior board item ever mentioned this.

A-07A-07 — Audit: when a player is removed, where do the opportunities go, in the simulator and in reality.done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item A-07, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Audit: when a player is removed, where do the opportunities go, in the simulator and in reality. | Reallocation error by position. FILES THE PLAN NAMES: research/playgrain/lineup/vacated_share.py MISSING: vacated_share.py No prior board item ever mentioned this.

E-06E-06 — Plant an effect of known size in the play data and require every gate to find it. Also the placedone

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item E-06, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Plant an effect of known size in the play data and require every gate to find it. Also the placebo battery (the astrology columns already exist as the negative control) and permutation nulls. Every gate publishes its power. FILES THE PLAN NAMES: research/playgrain/evalx/plant_controls.py MISSING: plant_controls.py No prior board item ever mentioned this.

E-07E-07 — The smallest detectable effect per layer, per window, game-clustered. It replaces the single 0.1done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item E-07, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The smallest detectable effect per layer, per window, game-clustered. It replaces the single 0.138 figure. FILES THE PLAN NAMES: research/playgrain/evalx/mde_table.py MISSING: mde_table.py No prior board item ever mentioned this.

E-09E-09 — Run `world-model-compiler-v1/cross-team-response-experiment.json`, which is `DESIGNED_NOT_RUN`: done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item E-09, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Run `world-model-compiler-v1/cross-team-response-experiment.json`, which is `DESIGNED_NOT_RUN`: G0 terminal Gaussian against the generator, five negative controls. FILES THE PLAN NAMES: research/playgrain/evalx/energy_experiment.py MISSING: energy_experiment.py No prior board item ever mentioned this.

F-06F-06 — Per-key hyperparameter search. `max_depth=3, l2=5.0` already fixes WR carries from −29.86% to −5done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item F-06, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Per-key hyperparameter search. `max_depth=3, l2=5.0` already fixes WR carries from −29.86% to −5.17%. | Board breaches cleared FILES THE PLAN NAMES: research/sweep_tree_params.py MISSING: sweep_tree_params.py No prior board item ever mentioned this.

F-08F-08 — Replace `intervals.SHIPPED_CURVE` with `sim.simulate` (`sim.py:813`), which has 0.7788 coverage done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item F-08, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Replace `intervals.SHIPPED_CURVE` with `sim.simulate` (`sim.py:813`), which has 0.7788 coverage and correct zero atoms. | Intervals from a generator, not a table FILES THE PLAN NAMES: none all present No prior board item ever mentioned this.

O-05O-05 — The throughput budget for S-14, measured on every release.done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item O-05, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The throughput budget for S-14, measured on every release. FILES THE PLAN NAMES: research/playgrain/operate/bench.py MISSING: bench.py No prior board item ever mentioned this.

O-06O-06 — New launchd jobs, all catch-up style.done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item O-06, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

New launchd jobs, all catch-up style. FILES THE PLAN NAMES: bin/*.plist MISSING: *.plist PRIOR EVIDENCE on the old board: C-178, C-163 REOPENED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): FALSE DONE. By its own result only 2 of 6 jobs were built (early-lines, playgrain-grade); playgrain-refresh and playgrain-issue do not exist and playgrain-loop is not loaded (launchctl list). The missing jobs are now N-10 (refresh) and N-04 (issue); playgrain-loop stays with L-07. Close this when N-10's and N-04's plists are loaded. UNBLOCKED 2026-09-25 07:40 MT (Claude): N-10 and N-04 are built and their plists are loaded (playgrain-refresh, playgrain-issue, playgrain-grade, playgrain-shadow in launchctl 07:38). Proceed.

O-07O-07 — The SQL findings stay, and the fabricated citations go. Nine of them independently corroborate tdone

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item O-07, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The SQL findings stay, and the fabricated citations go. Nine of them independently corroborate the credit-registry conservation defects. FILES THE PLAN NAMES: research/quarantine-20260919/, research/gemini-deep/ MISSING: , No prior board item ever mentioned this.

O-08O-08 — Remove the 60 worktrees whose branches have zero unmerged patches. The branches remain, so nothidone

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item O-08, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Remove the 60 worktrees whose branches have zero unmerged patches. The branches remain, so nothing is lost. FILES THE PLAN NAMES: bin/prune-worktrees.sh MISSING: prune-worktrees.sh No prior board item ever mentioned this.

P-08P-08 — 2026 in-season participation does not publish until after the season. Impute on-field sets from done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item P-08, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

2026 in-season participation does not publish until after the season. Impute on-field sets from weekly snap counts, depth charts, personnel tendencies and the pbp role ids (passer, rusher, target, tacklers, pass defenders). Trained and validated on 2016–2025, where truth is known. | Mean Jaccard ≥ 0.85 on hidden truth; error propagated as uncertainty into T-02. FILES THE PLAN NAMES: research/playgrain/participation_impute.py MISSING: participation_impute.py No prior board item ever mentioned this.

P-09P-09 — `game_context_asof`: roof, surface, elevation, travel_km, tz_shift, rest days, short week, officdone

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item P-09, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

`game_context_asof`: roof, surface, elevation, travel_km, tz_shift, rest days, short week, officials, weather ladder. All keyed by `valid_time`. | As-of tests. FILES THE PLAN NAMES: research/playgrain/load_context.py MISSING: load_context.py No prior board item ever mentioned this.

PW-104bin/janitor.sh not loaded in launchd, log stale since 2026-09-20 -- adjacent finding from PW-47, not fixeddone

Owner: Codex · Lane: — · Priority: 9

Why this matters

PW-47 (C-32, 2026-09-22 17:28 MDT), while fixing the identical bug for suite-check, found the SAME bug already live in janitor.sh and scoped it out on purpose rather than guess at a fix. Re-verified 2026-09-23 03:45 MDT while closing PW-47's own stale bookkeeping: the plist com.seano.playerweek.janitor.plist is present in ~/Library/LaunchAgents but `launchctl list` does not show it loaded, and data/logs/janitor.log's mtime is still 2026-09-20 21:20 -- unchanged, so the gap has held for three days, not just the two originally measured.

Evidence and next step

janitor.sh is not in bin/lane-watchdog.py's LANES table either, so its silence is invisible to the one system built to catch exactly this (bin/lane-watchdog.py:245's own comment: "'not loaded' reads as silence, not a fault" -- referring to this file by name as the precedent suite-check was built to not repeat). NOT A SIMPLE RELOAD: janitor.sh's own checks (research/plan_conditions.py, a six-condition "IS THE PLAN DONE" gate, and an LANES.md-era lane count) reference the pre-2026-09-22 board regime (the 174-item board construction.json's own `note` field says was replaced because it named only 21 of 119 plan items). Reloading it blind risks resurrecting a check built against a retired board rather than the current gate.py/construction.json regime. Whoever takes this needs to read research/plan_conditions.py first and decide: rewrite it against the current gate, retire it explicitly (remove the plist, log why), or confirm its checks still hold unmodified before reloading.

PW-105wh.hurdle_pred is a one-time 2019-2025 backtest, not a live sensor -- 2026 rows are zero and the box reads green anywaydone

Owner: Codex · Lane: — · Priority: 9

Why this matters

Found while closing PW-23: research/playgrain/build_hurdle.py's oof() and its --write path are hardcoded to test season <= 2025 (bh.oof(df, first=2019, last=2025)); wh.hurdle_pred holds 48,999 rows, all seasons 2019-2025, zero for 2026. research/playgrain/evalx/residual_ledger.py's own NEXT_SENSORS already named this gap ("the hurdle model's own P(active) output... this ledger's 'played' candidate uses the REALISED outcome, not the pre-kickoff forecast") before this probe confirmed it live.

Evidence and next step

buildsite/machine.py's 'hurdle model' box (A family, P(active) -> P(plays) -> share) reads `ev("{n:,} out-of-fold rows", "wh.hurdle_pred", need=1000)` -- a plain row-count threshold with no season filter, so it shows green on 48,999 historical rows while the table that would actually inform THIS season's decisions has none. Same shape as the hardcoded-green defects machine.py's own comments warn about (rapm_theta, caller_policy), just not one of the named ones yet. Measured 2026-09-23 (see PW-23's result): extending oof() to test season 2026 (training unchanged, still 2017-2025 only) and substituting that real forecast for bets against 2026 weeks 1-2 gave real_worth=-0.0334 against an oracle ceiling of +0.5288 -- the live forecast is not yet capturing the value the backtest promised, on a small n=689/2-week sample. Whoever takes this: (1) decide whether build_hurdle.py should run weekly with --write extended through the current season (an expanding-window refit each week, same pattern as build_playprob.py needs to be checked for), or (2) give the box a season-aware probe so it cannot read green on stale infrastructure -- both are real fixes, neither is done here. This item does not choose between them.

PW-110QB backup-conditional-level model, named as F-08's real remaining gap across seven follow-ups, never builtdone

Owner: Codex · Lane: — · Priority: 9

Why this matters

F-08 (construction.json, closed 2026-09-23) measured across seven follow-ups that sim.simulate() cannot match intervals.SHIPPED_CURVE's QB coverage (0.6923 vs nominal 0.80) because a benched starter's play-time uncertainty has no home in the simulator: a post-hoc floor patch and a per-draw conservation-preserving redistribution (sim.qb_draw_weights, built and proven in F-08's sixth follow-up) both help but neither closes the gap, because the real defect is one level up -- playprob.py's own module docstring names a backup QB's CONDITIONAL scoring level (his own rate when he actually plays) as measured too high, independent of how correctly his play probability is modeled. F-08 declined to fix this inside its own measurement script twice (fifth and seventh follow-ups), on the same reasoning both times: proxying around it via P(play) scores better out-of-sample but hides the level bug rather than fixing it, which is a worse trap than leaving the gap named.

Evidence and next step

Scope, if taken: fit a backup QB's expected fantasy output CONDITIONAL on actually playing (not blended with his near-zero bench rate) as its own model or feature, then feed that corrected level into sim.qb_draw_weights (already built, sim.py:258, proven by tests/test_qb_draw_weights.py) rather than the roster's raw historical share. Needs its own held-out measurement against real backup-QB games (injury-forced starts are the natural test set), not a re-run of F-08's coverage/correlation gate, which measures the simulator's structure and cannot see a level bug shared identically by both arms of an A/B. Not started -- no code exists for this anywhere in the codebase, research or production.

PW-156Q reader backlog at the approved schema-v2 claim pathnext

Owner: Codex · Lane: — · Priority: 9

Why this matters

PW-156 originally found an unscheduled local `pw read` step. That job was retired after the approved q01-daily-haiku reader took over schema-v2 claim production. The current question is whether eligible text is read within the reader's one-day bound after its owner pause ends. The Q pipeline gate now consumes the hold-aware backlog verdict, rather than treating a deliberate pause as a failure or reinstalling the retired job.

Dependency

q01-reader-owner-pause in data/holds.json since 2026-09-30; an explicit owner resume is required before reading resumes. Q experiments requiring fresh 2026 claims must declare the frozen information set while the pause holds.

Evidence and next step

The local `pw read` job was retired (5f600d76/ea8f7ba5); q01-daily-haiku writes the schema-v2 claims consumed downstream. Run `.venv/bin/python research/scientist/run/q_reader_backlog_check.py --db <snapshot>` for the reader's pending count and DAILY_LIMIT, and `.venv/bin/python tests/test_news_pipeline.py` for the Q pipeline gate that consumes its HELD/RED verdict. The check exits 0 GREEN, 1 RED without a hold above the bound, 2 UNREADABLE, or 3 HELD above the bound under the owner pause. Its planted no-hold control must fail. Preserve the owner pause; after an explicit resume, an over-bound RED needs bounded reader repair and a fresh live check.

PW-166Decouple QB from playermodel.py's shared cross-position stack, the path PW-110 named and did not attemptdone

Owner: Codex · Lane: — · Priority: 9

Why this matters

PW-110 (construction.json, closed 2026-09-28) tried five correction shapes for the backup-QB conditional-level bias playprob.py's docstring names; the one structural attempt (extending self.qb_share(d) to rushing) measured real conditional-level gains but failed whole-model held-out because PlayerModel.fit pools QB+RB+WR+TE into one shared 4-weight lstsq blend (b.raw(d) across every position) -- a QB-only stat_line change shifts that one shared vector and every other position's forecast with it (commits 11ff75a0, 0b50bd91, a34520df). PW-110 named but did not attempt decoupling the stack itself as out of its original scope.

Evidence and next step

Scope, if taken: give QB its own fit weights in PlayerModel.fit (playerweek/playermodel.py) instead of one shared regression across all four positions -- e.g. fit the stack per position group, or add a QB-specific term that does not enter the shared lstsq -- then re-attempt the qb_share(d)-on-cr rushing fix from PW-110 on top of it and re-run the same whole-model held-out check (run/pw110_wholemodel_check.py) to confirm RB/WR/TE no longer move. This changes core fitting behavior for every position, not just QB, so needs its own regression suite pass and a held-out whole-model MAE check on all four positions before any partial credit, not only the QB backup population PW-110 targeted.

PW-188Probe an untested limit (cannot) in commit:f0a39b2f3next

Owner: Claude · Lane: — · Priority: 9

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:f0a39b2f3: "Every row keeps its whole-run totals (for calibration and the top-10 list, where is the right question) and gains a parallel in_window_* set of the same columns; the incremental cache now stores each file's raw per-hour buckets (a pure function of the file) rather than the final rows, since the win" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-189Probe an untested limit (cannot) in commit:f6782038bnext

Owner: Claude · Lane: — · Priority: 9

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:f6782038b: "This fits the opposite direction: within each reset window, meter_pct(t) = Σ β_k · cumtokens_k(window, t) + offset_window + ε, via non-negative least squares (scipy.optimize.nnls — β can't be negative; offsets are free via the standard pos-minus-neg NNLS trick)." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-190Probe an untested limit (out-of-scope) in commit:8c97307cfnext

Owner: Claude · Lane: — · Priority: 9

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:8c97307cf: "H8 shown out of scope for this identity (row shuffle changes nothing)." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-191Probe an untested limit (out-of-scope) in commit:7d82e5022next

Owner: Claude · Lane: — · Priority: 9

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:7d82e5022: "Freeze H-270 as-of game-total preregistration (s-observation) Game layer, gametotal.predict(DEFAULT) baseline; team_week tempo + drive_game both-sides drive counts; never reads player_week, so H8 is out of scope; own-path order control declared." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-192Probe an untested limit (out-of-scope) in commit:09aab6867next

Owner: Claude · Lane: — · Priority: 9

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:09aab6867: "Market refine H-349 C2: H8 out of scope by reorder fixture; market-free Elo nil on bar Gradeable layer: player line (QB)." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-92168 of 353 test files are not green at baseline -- 114 error, 54 time out, unrelated to mutationdone

Owner: Codex · Lane: watchdog · janitor · Priority: 9

Why this matters

Found while re-verifying J-08's mutation harness (shard 0, 2026-09-23): tests/audit-gates-baseline.json now has full population coverage (353/353), and its 'skipped' bucket breaks down as 114 'baseline not green (rc=1)' + 54 'baseline timeout' -- 168 files, 48% of the suite, that the harness could not even mutation-test because the UNMUTATED test already fails or hangs. This is not a J-08 defect; J-08 correctly excludes what it cannot exercise. It is a finding about the suite itself: a test that cannot pass at baseline gives zero signal either way, mutated or not, and CI green on the suite as a whole says nothing about these 168.

Evidence and next step

No fix attempted here -- filing the finding, not the repair, per the rule that discovering a defect does not make it the discoverer's to work next. Next step for whoever picks this up: pull the 114 rc=1 filenames and their tracebacks (the harness's sandboxed run captures rc but this ledger entry format does not keep stderr -- would need a rerun with output captured, or grep test collection logs) and the 54 timeout filenames, and sort by whether the failure is a real regression (code moved under the test) versus an environment/fixture problem (missing fixture data, a since-renamed import, a live-service call that should have been in SAFETY_EXCLUDE). 12 additional files are fully green but 'cannot_fail' after 3 mutation attempts each -- listed on J-08's result -- which is a smaller, separate debt (the test runs fine, the harness just couldn't kill it).

Q-03Q-03 — `speaker_is` is unknown on 141 of 153 pressers, and it is the one field Q-36 turns on.done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item Q-03, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

`speaker_is` is unknown on 141 of 153 pressers, and it is the one field Q-36 turns on. FILES THE PLAN NAMES: playerweek/news.py all present No prior board item ever mentioned this.

Q-04Q-04 — Claims become as-of features for availability (A-02), role change (A-06) and play-caller intent done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item Q-04, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Claims become as-of features for availability (A-02), role change (A-06) and play-caller intent (B-02). Never for points directly: that was measured at +0.002 RMSE, where availability was −0.0038 Brier with a clean placebo. FILES THE PLAN NAMES: research/playgrain/sensors/claims_to_features.py MISSING: claims_to_features.py No prior board item ever mentioned this.

S-17S-17 — Any team-season against any team-season, retired players included. The north star's first demanddone

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item S-17, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Any team-season against any team-season, retired players included. The north star's first demand. | A football person reads the play-by-play without flinching. FILES THE PLAN NAMES: research/playgrain/hypothetical.py MISSING: hypothetical.py No prior board item ever mentioned this.

S-18S-18 — Whole seasons and playoffs. Ratings evolve between simulated weeks and injuries accrue through Adone

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item S-18, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Whole seasons and playoffs. Ratings evolve between simulated weeks and injuries accrue through A-04. | Division and playoff odds calibrated on 2004–2025. FILES THE PLAN NAMES: research/playgrain/season_sim.py MISSING: season_sim.py No prior board item ever mentioned this.

S-19S-19 — Compile the boosters (treelite or lleaves, or Numba lookup tables) if S-14 misses its budget. | done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item S-19, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Compile the boosters (treelite or lleaves, or Numba lookup tables) if S-14 misses its budget. | 16 games × 10,000 trajectories in under 60 minutes on the M5. FILES THE PLAN NAMES: research/playgrain/compile_models.py MISSING: compile_models.py No prior board item ever mentioned this.

T-08T-08 — Every rating family is scored on what it adds to the kernel and logged to the ledger. | —done

Owner: Codex · Lane: unassigned · Priority: 9

Why this matters

Plan item T-08, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Every rating family is scored on what it adds to the kernel and logged to the ledger. | — FILES THE PLAN NAMES: research/playgrain/talent/latent_eval.py MISSING: latent_eval.py No prior board item ever mentioned this.

PW-178Decide the external model cron cost acceptance clausedone

Owner: Sean · Lane: — · Priority: —

Dependency

PW-34 requires the actual token-costing 15-minute cron stopped or widened with before/after cost, or Sean explicit decline. External CTO-seat/Sean decision; no autonomous model work.

Evidence and next step

Record the external cron cost decision when Sean supplies it.

PW-179Define newsreader purposenext

Owner: Sean · Lane: — · Priority: —

Dependency

Sean: news reader currently useless; not throw tokens at it until we define what it should be doing

Evidence and next step

Discuss reader purpose with Sean; no autonomous model dispatch or reader restart.

PW-184Observe Anthropic budget admission with existing meterdone

Owner: Budget monitor · Lane: — · Priority: —

Dependency

Anthropic remains stopped: last shared meter reports 100% weekly used; preserve explicit owner outage hold.

Evidence and next step

Existing zero-model meter owns observations. Reading/reset/error receipt: ops/audits/dependency-budget-readback-20261001.json. Do not launch Anthropic calls or reset its hold.

PW-185Observe Gemini budget admission with existing meterdone

Owner: Budget monitor · Lane: — · Priority: —

Dependency

Explicit Gemini owner pause remains active. Live reading 2026-10-01T02:01:03.602358+00:00: weekly 65.9% and five-hour 100% remaining; prior 7% was an owner report, not the live meter.

Evidence and next step

Existing zero-model meter owns observation. Owner resumption is required by the preserved explicit hold, despite fresh fuel. Reading/reset/error receipt: ops/audits/dependency-budget-readback-20261001.json. No Gemini calls or hold reset.

PW-19Forecast did he play at all (worth +0.76 MAE as an oracle)done

Owner: Codex · Lane: — · Priority: —

Why this matters

Proposed by bin/residual-ledger.py, not by a person. The ledger feeds each candidate's REALISED value in as an oracle and records how much error disappears; that number is the value of measuring it perfectly, and ranking by it turns 'what should we build' from an opinion into a measurement. Sean, 2026-09-20, on being told the board was the one thing still written by hand: 'why not'.

Evidence and next step

MEASURED 2026-11-17, n=3760: knowing this perfectly takes our MAE from 4.3521 to 3.9382, worth +0.4139. That is a CEILING on what measuring it can buy, not a promise -- a variable worth 2 points as an oracle may be worth 0.2 once forecast. Sensor: the availability hurdle (C-53) -- P(active) before kickoff. [ledger:did he play at all]

PW-20Forecast his targets (worth +0.51 MAE as an oracle)done

Owner: Codex · Lane: — · Priority: —

Why this matters

Proposed by bin/residual-ledger.py, not by a person. The ledger feeds each candidate's REALISED value in as an oracle and records how much error disappears; that number is the value of measuring it perfectly, and ranking by it turns 'what should we build' from an opinion into a measurement. Sean, 2026-09-20, on being told the board was the one thing still written by hand: 'why not'.

Evidence and next step

MEASURED 2026-11-17, n=3760: knowing this perfectly takes our MAE from 4.3521 to 3.9125, worth +0.4396. That is a CEILING on what measuring it can buy, not a promise -- a variable worth 2 points as an oracle may be worth 0.2 once forecast. Sensor: the target policy: who gets the ball, given who is on the field. [ledger:his targets]

PW-21Forecast team rush attempts (worth +0.40 MAE as an oracle)done

Owner: Codex · Lane: — · Priority: —

Why this matters

Proposed by bin/residual-ledger.py, not by a person. The ledger feeds each candidate's REALISED value in as an oracle and records how much error disappears; that number is the value of measuring it perfectly, and ranking by it turns 'what should we build' from an opinion into a measurement. Sean, 2026-09-20, on being told the board was the one thing still written by hand: 'why not'.

Evidence and next step

MEASURED 2026-11-17, n=3760: knowing this perfectly takes our MAE from 4.3521 to 4.4341, worth -0.0820. That is a CEILING on what measuring it can buy, not a promise -- a variable worth 2 points as an oracle may be worth 0.2 once forecast. Sensor: team volume, rushing half. [ledger:team rush attempts]

PW-22Forecast team pass attempts (worth +0.36 MAE as an oracle)done

Owner: Codex · Lane: — · Priority: —

Why this matters

Proposed by bin/residual-ledger.py, not by a person. The ledger feeds each candidate's REALISED value in as an oracle and records how much error disappears; that number is the value of measuring it perfectly, and ranking by it turns 'what should we build' from an opinion into a measurement. Sean, 2026-09-20, on being told the board was the one thing still written by hand: 'why not'.

Evidence and next step

MEASURED 2026-11-17, n=3760: knowing this perfectly takes our MAE from 4.3521 to 4.4159, worth -0.0639. That is a CEILING on what measuring it can buy, not a promise -- a variable worth 2 points as an oracle may be worth 0.2 once forecast. Sensor: team volume: plays and pass rate given script and opponent. [ledger:team pass attempts]

PW-23Forecast his snap share (worth +0.33 MAE as an oracle)done

Owner: Codex · Lane: — · Priority: —

Why this matters

Proposed by bin/residual-ledger.py, not by a person. The ledger feeds each candidate's REALISED value in as an oracle and records how much error disappears; that number is the value of measuring it perfectly, and ranking by it turns 'what should we build' from an opinion into a measurement. Sean, 2026-09-20, on being told the board was the one thing still written by hand: 'why not'.

Evidence and next step

MEASURED 2026-11-17, n=3760: knowing this perfectly takes our MAE from 4.3521 to 3.6496, worth +0.7025. That is a CEILING on what measuring it can buy, not a promise -- a variable worth 2 points as an oracle may be worth 0.2 once forecast. Sensor: role forecasting from depth, usage and the injury ladder. [ledger:his snap share]

PW-24Forecast his carries (worth +0.32 MAE as an oracle)done

Owner: Codex · Lane: — · Priority: —

Why this matters

Proposed by bin/residual-ledger.py, not by a person. The ledger feeds each candidate's REALISED value in as an oracle and records how much error disappears; that number is the value of measuring it perfectly, and ranking by it turns 'what should we build' from an opinion into a measurement. Sean, 2026-09-20, on being told the board was the one thing still written by hand: 'why not'.

Evidence and next step

MEASURED 2026-11-17, n=3760: knowing this perfectly takes our MAE from 4.3521 to 4.1527, worth +0.1994. That is a CEILING on what measuring it can buy, not a promise -- a variable worth 2 points as an oracle may be worth 0.2 once forecast. Sensor: the carry policy and the backfield committee split. [ledger:his carries]

A-01A-01 — The Wed/Thu/Fri practice-status and game-status ladder, 2016–2026, as-of, from `inj_raw` and C-1done

Owner: Claude · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item A-01, lane ratings, lineups, policies, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The Wed/Thu/Fri practice-status and game-status ladder, 2016–2026, as-of, from `inj_raw` and C-14. | Every status row has a received-time or a reconstructed publish time. FILES THE PLAN NAMES: research/playgrain/lineup/inj_ladder.py MISSING: inj_ladder.py PRIOR EVIDENCE on the old board: C-75

B-01B-01 — Head coach, play-caller, OC and DC per team-week, as-of. `coordinators.csv` exists; fix `games_rdone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item B-01, lane ratings, lineups, policies, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Head coach, play-caller, OC and DC per team-week, as-of. `coordinators.csv` exists; fix `games_raw.home_coach`. FILES THE PLAN NAMES: research/playgrain/policy/coach_dim.py MISSING: coach_dim.py No prior board item ever mentioned this.

B-04B-04 — Go/punt/kick by coach. The late-and-close policy is its own model: kneel-downs, playing for the done

Owner: Claude · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item B-04, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Go/punt/kick by coach. The late-and-close policy is its own model: kneel-downs, playing for the field goal when tied, the two-point chart, timeout use, and the end-of-half "give up" mode both Gemini runs named as a collapse source. **This is where the missing mass on 3 comes from.** FILES THE PLAN NAMES: research/playgrain/policy/policy_fourth.py, policy/policy_endgame.py MISSING: policy_fourth.py, policy_endgame.py No prior board item ever mentioned this.

C-01C-01 — Per-source staleness measured by the *source's own timestamp*, not the job's exit code. It writedone

Owner: Claude · Lane: sensors · Priority: 0

Why this matters

Plan item C-01, lane sensors, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Per-source staleness measured by the *source's own timestamp*, not the job's exit code. It writes `alarms.jsonl` for the board, the only alarm path. Catches odds on 7 of 15 days, props 5 days stale, availability 48 h dark. FILES THE PLAN NAMES: research/playgrain/sensors/capture_health.py all present PRIOR EVIDENCE on the old board: C-01

C-02C-02 — Fix the `-3≤hours_to_kickoff≤6` window bug. Turn the eight one-shot captures into standing ones:done

Owner: Claude · Lane: sensors · Priority: 0

Why this matters

Plan item C-02, lane sensors, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Fix the `-3≤hours_to_kickoff≤6` window bug. Turn the eight one-shot captures into standing ones: `inj_snapshot` Wed/Thu/Fri, `avail_snap`, `presser`, `workload_claim` (where `lock_legal_wed` is TRUE on 0 of 715 rows). FILES THE PLAN NAMES: bin/institute-capture.sh, bin/institute-structured-capture.sh all present PRIOR EVIDENCE on the old board: C-02

C-03C-03 — Captures append to `data/land/*.parquet` and never touch DuckDB. One single-writer loader owns tdone

Owner: Claude · Lane: sensors · Priority: 0

Why this matters

Plan item C-03, lane sensors, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Captures append to `data/land/*.parquet` and never touch DuckDB. One single-writer loader owns the write lock, which ends `"status":"lost"`. FILES THE PLAN NAMES: research/playgrain/sensors/land_loader.py MISSING: land_loader.py PRIOR EVIDENCE on the old board: C-03

C-04C-04 — **The number you actually face.** Hourly from Monday 00:00 to Wednesday 23:59 MT: spreads, totaldone

Owner: Claude · Lane: sensors · Priority: 0

Why this matters

Plan item C-04, lane sensors, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

**The number you actually face.** Hourly from Monday 00:00 to Wednesday 23:59 MT: spreads, totals, moneylines, team totals and alt lines from every book in the feed. `williamhill_us` (Caesars, the operator of the Jackpot property) is flagged as the board proxy. `odds_history` already holds 32 books and 205 instants since 2023-09. Table: `early_lines`. FILES THE PLAN NAMES: research/playgrain/sensors/early_lines_capture.py MISSING: early_lines_capture.py PRIOR EVIDENCE on the old board: C-76, C-04

C-05C-05 — Props for all markets, not 6, hourly; de-vig method unchanged.done

Owner: Claude · Lane: sensors · Priority: 0

Why this matters

Plan item C-05, lane sensors, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Props for all markets, not 6, hourly; de-vig method unchanged. FILES THE PLAN NAMES: bin/odds-capture.sh, playerweek/props.py all present PRIOR EVIDENCE on the old board: C-05

D-02D-02 — One sitting, up to ~100 wagers that settle together. Portfolio Kelly over the joint trajectoriesdone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item D-02, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

One sitting, up to ~100 wagers that settle together. Portfolio Kelly over the joint trajectories, fractional, with every correlated position in the same game priced jointly. The maximum fraction at risk is read from `~/.config/playerweek/risk.json`, which no agent may write. FILES THE PLAN NAMES: research/playgrain/decide/stake_portfolio.py, playerweek/stake.py MISSING: stake_portfolio.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

D-03D-03 — For each wager, the worst line and price at which it is still a bet. The board will have moved bdone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item D-03, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

For each wager, the worst line and price at which it is still a bet. The board will have moved by the time you reach the window. FILES THE PLAN NAMES: research/playgrain/decide/walkaway.py MISSING: walkaway.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

D-04D-04 — Independent settlement from play-by-play into the append-only `data/proving.jsonl`. **Opened thidone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item D-04, lane kernel, drift, decision, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Independent settlement from play-by-play into the append-only `data/proving.jsonl`. **Opened this Tuesday with paper entries.** It cannot be backfilled, and it costs nothing. FILES THE PLAN NAMES: research/playgrain/decide/settle.py, decide/ledger.py MISSING: settle.py No prior board item ever mentioned this.

D-06D-06 — Closing-line value as a read-only diagnostic. The previous release keeps issuing in shadow, and done

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item D-06, lane kernel, drift, decision, due Fri 9 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Closing-line value as a read-only diagnostic. The previous release keeps issuing in shadow, and a CUSUM on paired CLV and Brier reverts the pointer. A seeded regression reverts in 3 weeks. FILES THE PLAN NAMES: research/playgrain/decide/clv_diag.py, loop/rollback.py MISSING: clv_diag.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

E-02E-02 — CRPS (margin, total, team totals, player lines), log loss (winner), energy score (joint: both QBdone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item E-02, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

CRPS (margin, total, team totals, player lines), log loss (winner), energy score (joint: both QBs plus both team scores), interval coverage. Confidence intervals by **game-block bootstrap**; paired Diebold–Mariano between releases. FILES THE PLAN NAMES: research/playgrain/evalx/scores.py MISSING: scores.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

E-03E-03 — Frozen-release backtest RED on 544 held-out gamesdone

Owner: Codex · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item E-03, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

MEASURED G4 RED on frozen S-16 release: raw early-price return -3.83% (95% interval -9.75% to +2.15%); affine candidate also RED. See result and release-specific full report. Evaluator-only, unreadable from the proposer's sandbox: skill against the closing line (nflverse `spread_line`/`total_line`, 1999+) and simulated return against the **early** ladder (the 540-snapshot backfill, T-21d to T-60m, six seasons, plus `early_lines`). Never a feature, never a training target. FILES THE PLAN NAMES: research/playgrain/evalx/market_diag.py MISSING: market_diag.py No prior board item ever mentioned this. REOPENED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): FALSE DONE. Leg A (spread -1.9%, total -3.8%) was scored on E-01's pre-c51817b2 predictions (totals ~88, INVALID). Leg B wrongly says the 540-snapshot early-line backfill is not built: data/land/odds-historical/ holds 540 JSONL files (2021-2026, rungs T21d-T60m) and odds_history holds 2023-25 (626/532/479 games, median lead 120h). RE-RUN: world-model predictions from E-01's re-run on the S-16 release (2021-2025 first); until those land use research/playgrain/eval-2024-fresh.csv and eval-2025-fresh.csv. Score against the EARLY ladder (snapshot nearest T-100h, median across books) AND the close; spread, total, moneyline; flat -110 ROI and hit rate with game-block bootstrap 95% intervals; mean CLV (close - early on the model's side); calibration slope; world model vs DYN, POWER, v0.1 and the rating model. Starting point: /tmp/claude-501/audit0924/grade_early.py and grade_close.py (2024 ATS vs early 48.9%, O/U 52.0%). OUTPUT: market_diag rows in data/results.duckdb + research/playgrain/evalx/g4_verdict.json with the G4 verdict (plan §6: return at early prices > 0 with a game-block interval excluding zero, mean CLV > 0, calibration slope 0.95-1.05) -- GREEN, or RED naming the failing condition.

E-04E-04 — One score per layer on the full issued population. The layers are availability, action, personnedone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item E-04, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

One score per layer on the full issued population. The layers are availability, action, personnel, target, outcome, magnitude, clock, drive, game and player line. Each faces the same frozen baselines at the same cutoff: carry-forward, pre-kickoff consensus, v1, the teamstate rival. A right total with a wrong workload story shows as two errors. FILES THE PLAN NAMES: research/playgrain/evalx/layer_scorecard.py MISSING: layer_scorecard.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

E-05E-05 — The Residual Attribution Ledger of §1. `oracle_pairs.py` tests candidate sensors **in pairs**, bdone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item E-05, lane ratings, lineups, policies, due Tue 6 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The Residual Attribution Ledger of §1. `oracle_pairs.py` tests candidate sensors **in pairs**, because `measurement_value.py` already proved greedy selection is blind: two sensors worth 4.4e-16 nats each were jointly worth 0.1766. FILES THE PLAN NAMES: research/playgrain/evalx/residual_ledger.py, evalx/oracle_pairs.py MISSING: residual_ledger.py, oracle_pairs.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

F-01F-01 — The measured fix. | MAE 4.2600 → 3.8748done

Owner: Claude · Lane: truth and the fantasy path · Priority: 0

Why this matters

Plan item F-01, lane truth and the fantasy path, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The measured fix. | MAE 4.2600 → 3.8748 FILES THE PLAN NAMES: none all present No prior board item ever mentioned this.

F-02F-02 — Give each QB a pass-attempt *share*, not the team's whole volume. | The QB gap of +0.975 closes done

Owner: Claude · Lane: truth and the fantasy path · Priority: 0

Why this matters

Plan item F-02, lane truth and the fantasy path, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Give each QB a pass-attempt *share*, not the team's whole volume. | The QB gap of +0.975 closes most of the way FILES THE PLAN NAMES: playermodel.py all present No prior board item ever mentioned this.

J-05J-05 — The durable results store, `data/results.duckdb`. It exposes `record(script, git_sha, populationdone

Owner: Claude · Lane: truth and the fantasy path · Priority: 0

Why this matters

Plan item J-05, lane truth and the fantasy path, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The durable results store, `data/results.duckdb`. It exposes `record(script, git_sha, population_sql, metric, value, ci_lo, ci_hi, n, seed, release)`. Nothing writes findings to `/tmp` again. | `tests/test_results_store.py`: a script run without a `record()` call fails CI. FILES THE PLAN NAMES: research/playgrain/results.py all present No prior board item ever mentioned this.

J-06J-06 — Score and fit on the unconditional population, with `E[pts] = P(play)·E[pts \| play]` and DNPs adone

Owner: Claude · Lane: truth and the fantasy path · Priority: 0

Why this matters

Plan item J-06, lane truth and the fantasy path, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Score and fit on the unconditional population, with `E[pts] = P(play)·E[pts \| play]` and DNPs as zeros. Include weeks 1–3, which have never been scored. | Every gate states its population string; `tests/test_gate_population.py`. FILES THE PLAN NAMES: playerweek/arms.py:229 all present No prior board item ever mentioned this.

J-07J-07 — Force `ORDER BY` on every sampled frame and hash the frame. This removes the ~0.006 pts/wk row-odone

Owner: Claude · Lane: truth and the fantasy path · Priority: 0

Why this matters

Plan item J-07, lane truth and the fantasy path, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Force `ORDER BY` on every sampled frame and hash the frame. This removes the ~0.006 pts/wk row-order floor that blocked 20 verdicts. | The same seed twice gives an identical hash. FILES THE PLAN NAMES: research/playgrain/evalx/determinism.py all present No prior board item ever mentioned this.

J-10J-10 — Delete the stale "irreducible" text; fail the build on `irreducible\|noise floor\|ceiling` applidone

Owner: Claude · Lane: truth and the fantasy path · Priority: 0

Why this matters

Plan item J-10, lane truth and the fantasy path, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Delete the stale "irreducible" text; fail the build on `irreducible\|noise floor\|ceiling` applied to football. The CORE-BRIEF's two sanctioned senses are allow-listed. | Green. FILES THE PLAN NAMES: tests/test_banned_words.py MISSING: test_banned_words.py No prior board item ever mentioned this.

L-07L-07 — The orchestrator: worst residuals → explanation brief → sandboxed proposal → evaluator → promotedone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item L-07, lane kernel, drift, decision, due Fri 9 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The orchestrator: worst residuals → explanation brief → sandboxed proposal → evaluator → promote or anti-library. Five proposals a night, hard-limited. The proposer sees a schema and a residual summary, never a row and never the holdout. FILES THE PLAN NAMES: research/playgrain/loop/nightly.py, bin/com.seano.playerweek.playgrain-loop.plist MISSING: nightly.py, com.seano.playerweek.playgrain-loop.plist PRIOR EVIDENCE on the old board: C-177 REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

L-11L-11 — It has 5 entries, all `UNRESOLVED`, while the docs call it the record of what is refuted. Populadone

Owner: Claude · Lane: unassigned · Priority: 0

Why this matters

Plan item L-11, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

It has 5 entries, all `UNRESOLVED`, while the docs call it the record of what is refuted. Populate it from the 48 rejected hypotheses, each with its population. Anything rejected on `played=1` is marked RETEST, not refuted. FILES THE PLAN NAMES: loop/anti-library.json all present No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

L-12L-12 — One row per packet: job, model, tokens, findings, confirmed, false positives, duplicates, downstdone

Owner: Claude · Lane: unassigned · Priority: 0

Why this matters

Plan item L-12, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

One row per packet: job, model, tokens, findings, confirmed, false positives, duplicates, downstream layer movement. Staffing follows measured yield. FILES THE PLAN NAMES: research/playgrain/loop/yield_ledger.py MISSING: yield_ledger.py No prior board item ever mentioned this. REPRIORITISED TO 0 on 2026-09-22 by Sean's instruction to cover the gaps. The objective is a SELF-IMPROVING machine and this item is on the chain that makes it one: L-07 is the loop, L-10 is the evaluator that stops the loop promoting a component gain that worsens the game forecast, E-* is how the machine tells whether anything is better, D-* is how it stakes. The plan's own order put L-07 on Fri 9 Oct -- the last day -- and L-10 unscheduled, which is exactly why zero proposals have ever been promoted. Measured at reprioritisation: L 0/6, E 1/10, D 1/7. 18 days to 10/10.

N-01N-01 — A first-class hold signal: a job that pauses says so, and every sensor reads itdone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Five times in 36 hours a sensor reported a designed pause as a fault -- build-serving deferring an archive, the odds job holding for budget, the watchdog restarting lanes for idling on an empty board, twice, and chain-report breaking a link on the pre-kickoff yield. Each was fixed as an exemption in the reader. The plan has no item for this because the plan assumed sensors would be right; the objective's 'limited human interaction' clause fails every time one is wrong and a person has to say so.

Evidence and next step

Added to the plan 2026-09-22 (revision noted in §6 of the plan document). Family N: what the objective needs that the design did not name.

N-02N-02 — The board IS the plan, enforced by a testdone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

The old board named 21 of the plan's 119 items and 87% of its own items came from elsewhere; nobody noticed for a week because nothing checked. A test that fails if any plan item is missing from the board, or any non-plan item lacks the PW- prefix, or a done item names a file that does not exist. Run by the gate. The confusion of 2026-09-22 cannot recur if it is a failing test rather than a thing a person notices.

Evidence and next step

Added to the plan 2026-09-22 (revision noted in §6 of the plan document). Family N: what the objective needs that the design did not name.

N-05N-05 — Stop bin/freeze-games.py rewriting 2026 game forecasts after kickoff; one canonical price per game per lockdone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Truth defect on the pricing path (build audit /tmp/claude-501/build-audit-2026-09-24.md §2). bin/freeze-games.py:229 runs `DELETE FROM game_bets WHERE season = 2026 AND model IN (?, 'teamstate')` and re-inserts every row with frozen_at = now. Measured 2026-09-24: every 2026 row, weeks 1-2 and TNF after kickoff included, carries frozen_at 2026-09-25 03:22:34 UTC, and TNF's pred_margin moved 5.63 -> 8.56. That breaks playerweek/gameserving.py:83 ('a forecast regenerated after kickoff is a memory'), and proving.jsonl and wagers took opposite sides on TNF.

Evidence and next step

BUILD: make game_bets append-only per (game_id, model, lock): insert only keys that are absent, never DELETE; skip and log any row whose kickoff <= now (kickoff is naive US/Eastern -- compare in that zone); add a `lock` column (playerweek.knowability.Lock) if missing, migrating existing rows in place without touching their frozen_at, and record that pre-fix 2026 frozen_at values are unreliable. Add one canonical view game_price_canonical(game_id, lock, model, pred_margin, pred_total, frozen_at) and point both decide-week.py/wagers and the proving ledger at it, so the two ledgers cannot take opposite sides again. ACCEPTANCE: tests/test_freeze_games_append_only.py on a temp DB: (1) running freeze-games twice leaves every existing row byte-identical, frozen_at included; (2) a game past kickoff is never inserted or changed; (3) re-planting the DELETE makes the test fail (seen red). Live: after the next scheduled freeze run, `SELECT count(*) FROM game_bets WHERE season=2026 AND frozen_at > kickoff` does not grow. Added to the plan 2026-09-24 (revision 6 of the plan document). Family N: what the objective needs that the design did not name.

O-03O-03 — The 38 duplicated 459 MB warehouses in `forecast-refresh/` become hard links to one content-addrdone

Owner: Claude · Lane: sensors · Priority: 0

Why this matters

Plan item O-03, lane sensors, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The 38 duplicated 459 MB warehouses in `forecast-refresh/` become hard links to one content-addressed copy, with manifests kept. 70 GB shrinks to a few. Enforce the written 21-day policy. FILES THE PLAN NAMES: bin/archive-retention.sh, research/playgrain/operate/dedupe_runs.py MISSING: dedupe_runs.py No prior board item ever mentioned this.

P-06P-06 — Load all 372 pbp columns into `plays_wide`; `plays` stays as a stable view. Extend the banned lidone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item P-06, lane ratings, lineups, policies, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Load all 372 pbp columns into `plays_wide`; `plays` stays as a stable view. Extend the banned list to `spread_line`, `total_line`, `vegas_*`, `*_vegas`, `odds*`. | `test_warehouse.py` extended: 0 market columns reachable from `plays_wide`. FILES THE PLAN NAMES: research/playgrain/load_plays_wide.py, schema.py all present PRIOR EVIDENCE on the old board: C-79, C-100

P-07P-07 — Participation 2016–2025 → `participation(game_id, play_id, gsis_id, side, pos)` and `play_persondone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item P-07, lane ratings, lineups, policies, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Participation 2016–2025 → `participation(game_id, play_id, gsis_id, side, pos)` and `play_personnel(offense_personnel, defense_personnel, defenders_in_box, n_pass_rushers, formation, route, man_zone, coverage_type, time_to_throw, was_pressure)`. **Every charted post-snap field is tagged `POST_SNAP` in the contract** and usable only as a training target or a rating input, never as a same-play feature. This is the Gemini B-arm's own correction about `defense_coverage_type`. | Leak witness green. FILES THE PLAN NAMES: research/playgrain/load_participation.py MISSING: load_participation.py No prior board item ever mentioned this.

P-10P-10 — One identity: GSIS canonical through `dp_ids` (35 id systems, 100% populated) in place of `xwalkdone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 0

Why this matters

Plan item P-10, lane ratings, lineups, policies, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

One identity: GSIS canonical through `dp_ids` (35 id systems, 100% populated) in place of `xwalk` (29% null). Kickers and DST included; all kickoffs in UTC with the Eastern source column kept. Fixes the week-1 FantasyPros-id join that "silently matches nothing". | 0 orphan ids; `SCHEDULE_IDENTITY_CONFLICT` (`player_scorecard.py:731-750`) re-keyed from exact-UTC equality to `game_id`. FILES THE PLAN NAMES: research/playgrain/players_dim.py MISSING: players_dim.py No prior board item ever mentioned this.

PW-121Watchdog must read log-record clocks rather than timestamps inside quoted outputdone

Owner: Codex · Lane: operations · Priority: 0

Why this matters

Worker 1 entered with the gate closed: all lanes alive and watched reported 7 dead. This unit was restricted to that condition.

Evidence and next step

bin/lane-watchdog.py last_evidence searched anywhere in a line for a timestamp. A JSON-quoted tool result containing a build evidence line could replace the actual wrapper clock with an older or future value. Require the timestamp at the beginning of the record, matching last_start. Preserve the existing filesystem clock for unstamped snapshot logs. This narrows embedded-payload matching; plain replayed lines and the existing mtime fallback are unchanged.

PW-129Build provider holds must reach the watchdog registrydone

Owner: Codex · Lane: operations · Priority: 0

Why this matters

Gate condition all lanes alive and watched blocked Q-01; provider waits logged HOLD but omitted the hold registry.

PW-170Tests must never open the production warehouse; Tuesday issuance gets a reserved lock windowdone

Owner: Claude · Lane: — · Priority: 0

Why this matters

Incident #7: suite-check.py ran tests against data/playerweek.duckdb and held its lock for 3 h+ on Tue 09-29, failing week-4 issuance twice (0/16 priced) and the 04:00 refresh. suite-check is DISABLED until this lands.

Evidence and next step

(1) The test suite reads a snapshot copy of the warehouse (or a fixture), never the live file; a guard makes a test that opens the live path fail. (2) Tue 03:30-12:00 MT is a reserved window: non-issuance jobs that need the write lock yield (holds.json) and issuance/refresh get priority. (3) Re-enable com.seano.playerweek.suite-check only after (1). ACCEPTANCE: suite run leaves lsof on the live DB empty; a planted long lock holder during the window yields; next Tuesday issuance prices 16/16 on the first attempt.

PW-171arm-forecast-refresh holds the warehouse write lock for its whole rebuild (44+ min) and is killed and retried in a loopdone

Owner: Claude · Lane: — · Priority: 0

Why this matters

Incident #13 (ops/incidents/INCIDENT-LOG.md): on Tue 09-29 arms.rebuild_forecast held data/playerweek.duckdb for 44 min at 50% CPU after being killed (rc=143) at 06:54, 07:28 and 09:06, each retry re-taking the lock. The hourly chain could not finish, so the build gate stayed closed all morning. It normally takes about 5 min (09-22). It is DISABLED (launchctl disable com.seano.playerweek.arm-forecast-refresh) until this lands.

Evidence and next step

Compute the arm forecasts from a read-only connection (or a snapshot) with no lock held, then write them in one short write transaction, well under a minute. Find why this run took 44+ min against about 5 min on 09-22. Give the job a wrapper timeout that releases cleanly and does not retry while the chain or the Tuesday issuance window needs the lock. ACCEPTANCE: a full rebuild holds the write lock under 60 s (measured via lsof/timestamps); it runs during a live chain without a LockTimeout in chain-run.log; then re-enable the launchd job and verify one scheduled run.

PW-32Week totals withheld while their own component total is publisheddone

Owner: Claude · Lane: — · Priority: 0

Why this matters

Sean found this, which means nothing else did. Weeks 1 and 2 read 'Not published -- No stored product was supplied' for my_total while group_totals.mine.starters.projection sat beside them at 280.20 and 273.65. The new check finds EIGHT such weeks, not two.

Evidence and next step

THE CHECK EXISTED AND COULD NOT SEE IT. validate-payload.py asks whether a GROUP total is missing while its rows are present. The group total was present; the thing missing was one level up, the WEEK total, and nothing looked there. A validator that covers every level but one inspires exactly the confidence that makes the gap expensive. The check is now in and proven red against the live payload -- 8 gaps: weeks 1, 2 and 15 among them, both sides. WHAT IS LEFT: the CAUSE. bin/build-serving.py reads w.get('my_total') from upstream, so the producer's source already withholds it. Weeks 1 and 2 are PLAYED and week 15 is not, so 'played weeks differ' is not the whole story -- find what actually distinguishes the eight, and fix it there rather than defaulting the total to the group sum at render time. A number computed in the publisher to paper over a missing product is the same defect wearing arithmetic.

PW-57Grade the GAME SHAPE — the rung above the statsdone

Owner: Claude · Lane: s-observation · Priority: 0

Why this matters

Sean, 2026-09-22: "the goal is that the play by play, the shape of the game can be graded, not just the stats."

Evidence and next step

The ladder is play sequence -> drives -> team shape -> stat line -> points, and only the bottom rung is graded. drive_week already exists (1999-2025: start, end, points, clock burned, game state) and sim.py generates game scripts, so the drive rung is gradeable TODAY. research/aar_gameshape.py grades team totals for a single week and is not part of the standing review. X-01 and X-03 did the play-level comparison once as research and found real and simulated states DISJOINT, not shifted — that finding never became a standing grade. An error inherited from the team rung is not an error born at the player rung, and only this separation tells you which.

PW-61Codex is a peer model and it has been dark since 2026-09-17done

Owner: Claude · Lane: operations · Priority: 0

Why this matters

80 codex/* branches, a passdown written BY Codex TO Claude, and no lane, plist, board item or plan family mentions it. Its charter is the systems scientist: architecture contracts, promotion gates, decisive experiment design, leakage audit, and independent red-team before real-money use. That is precisely the role nothing currently fills. Decide the runner, the clock and the mandate, and put it on launchd like every other lane. See docs/WHAT-I-WAS-MISSING.md §1.

PW-73Three build lanes were busy and the board showed nothing in flightdone

Owner: Claude · Lane: construction · Priority: 0

Why this matters

Sean: "only mice are working?" — they were not; three build-loop processes and two witness leases were running at that moment.

Evidence and next step

The plan map only colours PLAN items, and the build lanes mostly work PW- items, which are self-generated and deliberately excluded from it. So the one section he reads to answer 'what is getting worked' is structurally blind to most of the work. The In-flight card carries it, but it sits below the map and reads as a list rather than as the answer. Make the answer visible where the question is asked: every lane, what it holds, plan item or not.

PW-76The fuel gate reads 62% while the client reads 3% — a safety device off by 20xdone

Owner: Claude · Lane: operations · Priority: 0

Why this matters

Sean, 2026-09-22: "Session usage at 67%, weekly usage at 3%." playerweek.cli fuel said "weekly 62% of 100% limit" at the same moment.

Evidence and next step

Every lane gates on this number: a FUEL STOP halts the build lanes, the AAR, the scientists and the referee. If the true weekly figure is 3% and the gate believes 62%, the machine has been throttling itself against a wall that is not there — all night, while three providers were being asked to cover each other. fuel.py reads seven_day.utilization from api.anthropic.com/api/oauth/usage. Probed just now the endpoint returned HTTP 429, so `fuel` was serving its CACHE while reporting "source: live meter" — which is its own defect: a cached reading must not describe itself as live. TWO THINGS TO SETTLE: whether seven_day is the same bucket the client calls "weekly" (there may be separate Opus and overall buckets), and why the cache is labelled live. Until then treat 62% as unverified — do NOT widen any throttle on the strength of it, because the failure mode of being wrong in the other direction is spending his money. CORRECTION 2026-09-22 21:15: Sean — "Nothing ran dry today." He is right about his subscription and I was wrong to say I burned it. The agy CLI is signed in as his Ultra account and still returns "Individual quota reached… upgrade your subscription"; the CLI has its own per-account ceiling, and that is what the mice emptied. His Deep Think and Deep Research capacity was never touched. This also means the 3%/62% gap above is a question about ANTHROPIC's buckets only — the Gemini story is a separate tank entirely.

S-06S-06 — Calibration measured: held-out reliability FAILSdone

Owner: Codex · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item S-06, lane kernel, drift, decision, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

MEASURED NEGATIVE RESULT on frozen S-16 release; not promoted to live prices. 2024/2025 isotonic home-win, cover and over slopes miss 0.95–1.05; see result and release-specific candidate artifact. Isotonic calibration on a rolling holdout. The current model says 0.606 where 0.759 happens. Unbuilt since revision 2. | Reliability slope 0.95–1.05. FILES THE PLAN NAMES: research/playgrain/calibrate.py all present No prior board item ever mentioned this. REOPENED 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): FALSE DONE. Fitted on the broken pre-c51817b2 2008/2009 origins plus eval_sim; its own result says the gate FAILS and cross-era calibration hurts (research/playgrain/calibrate.json). REDO WITHIN ONE ERA on E-01's re-run (S-16 release): fit on 2023 -> apply 2024, fit <=2024 -> apply 2025. Calibrate P(home win), P(cover) at the posted spread including the -3/-7 key numbers, P(over), and the margin sd (12.6 simulated vs 14.5 real, research/playgrain/drift/x05_score_shape_RESULT.md). Until E-01's 2023-2025 origins land, build and test on research/playgrain/eval-2024-fresh.csv / eval-2025-fresh.csv and rerun when they do. OUTPUT: calibrate.json with reliability slope per market and season, plus the calibration map E-03 and N-04 apply. ACCEPTANCE: reliability slope 0.95-1.05 on the held-out season, or the failure named.

S-08S-08 — Early stopping on a validation season, not an inference cap at 60 rounds. 115 integer yard bins.done

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item S-08, lane kernel, drift, decision, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Early stopping on a validation season, not an inference cap at 60 rounds. 115 integer yard bins. Passes split into air yards and YAC (`mag_air`, `mag_yac`). Every booster registered with its training hash. | OOS log-loss monotone in rounds. FILES THE PLAN NAMES: fit_models.py all present No prior board item ever mentioned this.

S-10S-10 — Semi-Markov durations: the elapsed-time *distribution* per play type, state, tempo and coach, indone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item S-10, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Semi-Markov durations: the elapsed-time *distribution* per play type, state, tempo and coach, in place of five constants. Two-minute warning, runoffs, timeouts. | Plays per game 152.3 ± 2%. FILES THE PLAN NAMES: research/playgrain/kernel/clock.py MISSING: clock.py No prior board item ever mentioned this.

V-01V-01 — The rendered page as a standing surfacedone

Owner: Codex · Lane: s-observation · Priority: 0

Why this matters

Every other check validates what we PRODUCE — the payload, the serving columns, the 129-link data chain — and every one was green while week 1 of the consumer site was visibly wrong in a browser. The rules here are comparative because that is what a row-by-row validator structurally cannot be: a settled week whose projection equals its actual, a week rendering far below its siblings' median. Runs every 30 minutes at zero model tokens; findings become crumbs the mice claim and fix, and appear on the build board. An audit that could not open the site writes ok:false and becomes its own crumb, because no findings reads exactly like a clean site.

X-01X-01 — **First thing built.** k-step rollouts from real states (k = 1, 2, 4, 8, 16, to end of drive); ddone

Owner: Claude · Lane: kernel, drift, decision · Priority: 0

Why this matters

Plan item X-01, lane kernel, drift, decision, due Tue 22 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

**First thing built.** k-step rollouts from real states (k = 1, 2, 4, 8, 16, to end of drive); divergence curves for yards per play, completion and field position, stratified by the game's realised-minus-rated quality. It confirms or kills the §3 diagnosis. | The result decides the order of X-02/03/04. FILES THE PLAN NAMES: research/playgrain/drift/drift_decompose.py MISSING: drift_decompose.py No prior board item ever mentioned this.

C-06C-06 — Fix the Eastern-into-UTC index bug. Capture the forecast ladder (lead 7→0 days, hourly) for everdone

Owner: Claude · Lane: sensors · Priority: 1

Why this matters

Plan item C-06, lane sensors, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Fix the Eastern-into-UTC index bug. Capture the forecast ladder (lead 7→0 days, hourly) for every venue. Backfill historical *forecasts*, not observations, from Open-Meteo's historical-forecast archive. Remove `temp_observed`/`wind_observed` from the five feature tables that leak them. FILES THE PLAN NAMES: research/playgrain/sensors/weather_asof.py, playerweek/openmeteo.py:40-59 MISSING: weather_asof.py No prior board item ever mentioned this.

C-07C-07 — Referee crew per game, historical from nflverse `officials` plus the weekly assignment release.done

Owner: Claude · Lane: sensors · Priority: 1

Why this matters

Plan item C-07, lane sensors, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Referee crew per game, historical from nflverse `officials` plus the weekly assignment release. FILES THE PLAN NAMES: research/playgrain/sensors/officials_capture.py MISSING: officials_capture.py PRIOR EVIDENCE on the old board: C-07

C-09C-09 — Daily roster diff: signings, IR, practice-squad elevations (the Saturday tell), waiver claims.done

Owner: Claude · Lane: sensors · Priority: 1

Why this matters

Plan item C-09, lane sensors, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Daily roster diff: signings, IR, practice-squad elevations (the Saturday tell), waiver claims. FILES THE PLAN NAMES: research/playgrain/sensors/transactions_capture.py MISSING: transactions_capture.py PRIOR EVIDENCE on the old board: C-09

C-11C-11 — One refresher for everything on the nflverse release path: pbp, participation, FTN charting, NGSdone

Owner: Claude · Lane: sensors · Priority: 1

Why this matters

Plan item C-11, lane sensors, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

One refresher for everything on the nflverse release path: pbp, participation, FTN charting, NGS (2026 rows are missing), PFR advanced stats (2026 missing), snaps, depth charts, rosters_weekly, injuries, officials, contracts. Records `fetched_at` and the upstream `last_modified`. FILES THE PLAN NAMES: research/playgrain/sensors/nflverse_refresh.py MISSING: nflverse_refresh.py No prior board item ever mentioned this.

C-12C-12 — Parse the FTN payloads already sitting in `evidence.sqlite` into `ftn_charting`: play-action, scdone

Owner: Claude · Lane: sensors · Priority: 1

Why this matters

Plan item C-12, lane sensors, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Parse the FTN payloads already sitting in `evidence.sqlite` into `ftn_charting`: play-action, screen, RPO, motion, blitzers, rushers, out-of-pocket, catchable, contested, drop, created reception, read thrown, QB location, hash. FILES THE PLAN NAMES: research/playgrain/sensors/ftn_parse.py MISSING: ftn_parse.py PRIOR EVIDENCE on the old board: C-77

F-03F-03 — Train on `played=0` rows too, so the model *can* emit a low number. | Re-enables `loss="absolutedone

Owner: Claude · Lane: truth and the fantasy path · Priority: 1

Why this matters

Plan item F-03, lane truth and the fantasy path, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Train on `played=0` rows too, so the model *can* emit a low number. | Re-enables `loss="absolute_error"`, which passed every gate and was silently killed FILES THE PLAN NAMES: none all present No prior board item ever mentioned this.

J-01J-01 — Remove `mkt_total`, `mkt_spread`, `mkt_team_total` from `team_volume` and refit. At serve time tdone

Owner: Claude · Lane: truth and the fantasy path · Priority: 1

Why this matters

Plan item J-01, lane truth and the fantasy path, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Remove `mkt_total`, `mkt_spread`, `mkt_team_total` from `team_volume` and refit. At serve time they are constants that collapse into the intercept, so they buy nothing and contaminate every backtest. Call `knowability.assert_clean` *inside* `lines_for`; it never runs on the model path today. | `tests/test_no_market_reachable.py` fails if any market-derived column is reachable from `ALL_FEATURES` or `team_volume`. FILES THE PLAN NAMES: playerweek/playermodel.py:95-170, playerweek/share.py all present No prior board item ever mentioned this.

J-02J-02 — Remove the `HOME_FIELD_MARGIN/2` double count (1.748 pts/game). Home field is derived from scheddone

Owner: Claude · Lane: truth and the fantasy path · Priority: 1

Why this matters

Plan item J-02, lane truth and the fantasy path, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Remove the `HOME_FIELD_MARGIN/2` double count (1.748 pts/game). Home field is derived from schedule and venue only. | Mean home residual on the 2019–2025 holdout within ±0.2. FILES THE PLAN NAMES: playerweek/gameproj.py all present No prior board item ever mentioned this.

J-03J-03 — Re-run every headline accuracy number on clean features and the full population; publish before/done

Owner: Claude · Lane: truth and the fantasy path · Priority: 1

Why this matters

Plan item J-03, lane truth and the fantasy path, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Re-run every headline accuracy number on clean features and the full population; publish before/after. Expect the numbers to get worse. | Table written to the results store. FILES THE PLAN NAMES: research/playgrain/evalx/remeasure_ledger.py MISSING: remeasure_ledger.py No prior board item ever mentioned this.

J-08J-08 — Mutation harness: inject a known defect per test and require a red result. 73 of 156 tests curredone

Owner: Claude · Lane: truth and the fantasy path · Priority: 1

Why this matters

Plan item J-08, lane truth and the fantasy path, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Mutation harness: inject a known defect per test and require a red result. 73 of 156 tests currently cannot fail. | 0 un-failable tests; the auto-shipper's two first. FILES THE PLAN NAMES: tests/audit_gates_can_fail.py MISSING: audit_gates_can_fail.py No prior board item ever mentioned this.

J-09J-09 — Cross-fit every calibration, isotonic and stack weight that is in-sample today (D-232: 0.9984 indone

Owner: Claude · Lane: truth and the fantasy path · Priority: 1

Why this matters

Plan item J-09, lane truth and the fantasy path, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Cross-fit every calibration, isotonic and stack weight that is in-sample today (D-232: 0.9984 in-sample against 1.1383 held out). | Held-out `got/said` within 0.97–1.03 by position. FILES THE PLAN NAMES: none all present No prior board item ever mentioned this.

J-11J-11 — One full trace: reproduce a sealed issuance bit for bitdone

Owner: Claude · Lane: unassigned · Priority: 1

Why this matters

Gemini 1/10: the strongest audit of S-16 we lack; also catches silent nondeterminism.

Evidence and next step

One full trace (Gemini 1/10): pick one issued game and one player; re-run the sealed S-16 issuance from its release and recorded seed and reproduce the trajectories bit for bit; prove player stats sum to team stats on a sampled trajectory; show the decision row was written before kickoff and never updated; match the served payload hash to the sealed row. ACCEPTANCE: The test passes on a real issuance and fails when any link is broken (seen red).

N-11N-11 — What the Wednesday line misses: screen Tuesday-knowable inputs against (close − early) and (actual − early), 2023–25done

Owner: Claude · Lane: kernel, drift, decision · Priority: 1

Why this matters

The actual search for an edge (build audit /tmp/claude-501/build-audit-2026-09-24.md §4, §6). Verdict RED: no engine beats the early or closing number (world model ATS vs early 48.9% / 49.2% in 2024/2025; the line does not move toward the model, -0.06 / -0.02 pts). Nothing on the board asks what information the Wednesday number does not already contain. The data is on disk, so this is cheap.

Evidence and next step

INPUTS: early number = snapshot nearest T-100h, median across books, from odds_history (2023-25: 626/532/479 games, median lead 120h) and data/land/odds-historical/ (540 JSONL, 2021-2026, rungs T21d-T60m) -- same rule as /tmp/claude-501/audit0924/grade_early.py; close + results from data/cache/games.csv. Candidate inputs knowable by Tuesday: the A-03 injury-status forecast, QB change vs prior start, forecast wind/precip at issue time, rest days and travel distance, and (world model - early) from E-01 on the S-16 release. BUILD research/playgrain/evalx/edge_screen.py. Pre-register the input list and every threshold in the file header BEFORE reading any result. For each input: regress (close - early), (actual margin - early spread) and (actual total - early total) on it; game-block bootstrap 95% intervals; per-season sign; hit rate and ROI at -110 of the rule 'bet when the input says X'; Holm-adjusted p across inputs. OUTPUT: research/playgrain/evalx/edge_screen.json + a results row (script edge_screen.py) in data/results.duckdb. ACCEPTANCE: every pre-registered input reported, nulls included, with n, interval and per-season sign, and one verdict line naming each input whose interval excludes zero in all three seasons -- or 'none'. Added to the plan 2026-09-24 (revision 6 of the plan document). Family N: what the objective needs that the design did not name.

PW-112Probe an untested limit (cannot) in commit:26e0e681done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:26e0e681: "Chose E-10 as shard 2s remainder cannot be settled, S-18 is built." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-146N-04 PlayerLedger allocates on raw trailing-3 shares; the Kalman centre is measured better held outdone

Owner: build-w2 · Lane: unassigned · Priority: 1

Why this matters

prospective_shadow.roster_from_history passes target_share_l3/carry_share_l3 with kappa=None. A player with one or two games feeds a raw 1-2 game mean. rating_posterior.json shows the Kalman centre beating l3 by +1.66 [1.24,2.08] and +1.53 [1.19,1.89] LL per team-game on 2024/25 targets, and +1.49/+0.96 on carries.

Evidence and next step

Wire the Kalman-centre (and fitted Dirichlet kappa) into the issuer's roster, the same path sim2.load_roster(posterior=...) already supports. The change touches issuance code: do it as a new release id via kernel_release.py, never an edit to the pinned k51f week-4 issuance. ACCEPTANCE: issuer rosters use a shrunk share centre; held-out player-line log-lik vs l3 recorded on 2024/25.

PW-151B and Q contribution on the status page is frozen at 2026 weeks 1-2: re-run the with/without-arm refits every weekdone

Owner: Codex · Lane: — · Priority: 1

Why this matters

Sean's status page shows how the B (policies) and Q (language) arms move prediction accuracy. The only with/without comparison is a set of refits built 2026-09-20 covering weeks 1-2, so the numbers never update.

Evidence and next step

Direct current status-metrics proof: arm-live weeks 1–3; B ff.points -1.515, Q ff.points +0.221, n=1157. Installed wrapper last rc0 2026-09-27 20:26 MT; dispatcher budget corrected to 14,400s. Evidence: docs/evidence/PW-151-2026-09-29-present-proof.json.

PW-157schedule.JOBS is out of sync with launchd: 52 loaded jobs it doesn't declare, 50 declared jobs not loadeddone

Owner: build-w5 · Lane: — · Priority: 1

Why this matters

Found while triaging PW-92's 168 baseline-not-green files: tests/test_schedule_clock.py measured (2026-09-27) 52 launchd jobs running with no schedule.JOBS() entry -- uninstall() cannot remove them and plan() does not cost them -- and 50 declared jobs that are NOT currently loaded (`pw schedule --install --confirm` would load them). Separately, 15 of 47 clock jobs fire more than 5 minutes from their declared time (median +3.7min, worst +60.4min on one brief job) -- the same launchd-calendar-drift class as the 2026-09-25 TZ fix, measured fresh and still nonzero after it.

Evidence and next step

playerweek/schedule.py's JOBS registry needs reconciling against `launchctl list` output in both directions (extra loaded jobs, missing loaded jobs), and the >5min clock-drift jobs need the calendar-agent restart or reinstall this project has used before. Full job lists are in data/pw92-rc1-recheck.json's test_schedule_clock.py tail; do not re-run just to confirm it is still drifting, this is already a fresh measurement.

PW-159tests/test_aar.py crashes unhandled before its own PW-56 check ever runsdone

Owner: build-w7 · Lane: — · Priority: 1

Why this matters

Found 2026-09-27 while verifying PW-56's per-stat grading: tests/test_aar.py section 3 (the head_to_head common-sample check) raises an unhandled aar.EmptyComparison on the live warehouse -- 'playermodel-live', a partial 2026 source, shares no common player-week with the seven fully-backfilled sources under arm SBQ, so head_to_head(c, srcs) over ALL of them is empty by construction, not by regression. The script has no try/except around that call, so it crashes before reaching section 6 (PW-56's own receiving_yards join check) or printing any PASS/FAIL line. Reproduced on unmodified HEAD, so pre-existing and not introduced by this session's edits.

Evidence and next step

tests/test_aar.py line ~49: either wrap the head_to_head(c, srcs) call in the same try/except EmptyComparison pattern already used lower in the file for the impossible-filter case (record it as informational, not FAIL, since a partial live source lacking full-history overlap is expected), or exclude '-live' sources from the ALL-sources common-sample check and cover them with their own narrower comparison. Verify by re-running tests/test_aar.py end to end and confirming it reaches a FAIL/PASS summary line even while playermodel-live exists.

PW-173Probe an untested limit (cannot) in commit:6bf5ddf5done

Owner: Codex · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:6bf5ddf5: "No code change this unit: the locking fix already landed and is unaffected; the remaining acceptance gap (a completed end-to-end run's held_s) cannot be produced without either waiting out pid 10699 past the 20-minute alarm or starting a duplicate run, both against standing instructions." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-204Probe an untested limit (not-worth) in commit:048ba61a3done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:048ba61a3: "H-522's J is real at P.epa, not worth points." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-206Probe an untested limit (banned-floor) in commit:206276b25done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:206276b25: "THE MEASUREMENT, the test's own (tdoe.measure, run on a pinned snapshot 2026-10-15): arm SBQ fitted on season <= 2023 and scored on the 15,724 identical played, rostered 2024+ player-weeks, both arms fitted in one process, a third identical fit as the noise floor." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-27Probe an untested limit (cannot) in board:C-14done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from board:C-14: "The injury-report ladder and depth charts "cannot be walked backwards" only through one endpoint." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work. ID RE-PREFIXED 2026-09-22. This item was filed as C-27, which collides with the plan's own C-01..C-18 namespace. A lane reading 'C-19' could not tell a residual-ledger proposal from a plan item, and the two mean entirely different things. Locally-generated items carry PW- now; the plan owns the bare letter prefixes.

PW-30Probe an untested limit (not-worth) in commit:d81a4a26done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:d81a4a26: "Team play-calling tendency not worth carrying as a feature." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work. ID RE-PREFIXED 2026-09-22. This item was filed as C-30, which collides with the plan's own C-01..C-18 namespace. A lane reading 'C-19' could not tell a residual-ledger proposal from a plan item, and the two mean entirely different things. Locally-generated items carry PW- now; the plan owns the bare letter prefixes.

PW-33A first-class 'deliberate hold' signal, instead of case-by-case exemptionsdone

Owner: Claude · Lane: — · Priority: 1

Why this matters

Audit finding. Five separate times in 36 hours a sensor reported a designed pause as a fault: build-serving exiting 1 on a deferred archive, the odds sensor calling a budget-driven hold STALE, the watchdog restarting build lanes for idling on an empty board, then the mice for the same, and chain-report breaking a link on the pre-kickoff yield. Each fix was correct and each was a one-off exemption written into the reader.

Evidence and next step

That is the wrong layer. A job that decides to hold KNOWS it is holding; the reader is guessing. Every lane should write its hold reason to one place (data/holds.json: job, since, reason, expected-until) and every sensor should consult it before calling a silence a fault. Then a sixth deliberate hold does not need a sixth exemption. The five existing exemptions become the first five entries, and stay as the test that the mechanism catches what they caught.

PW-35AAR: variance error is 58% of our total, and it is calibration's layerdone

Owner: Claude · Lane: calibration · Priority: 1

Why this matters

The after-action review attributes 58% of all forecast error we have actually made to variance — 20,655 rows, mean |error| 2.83. That layer belongs to the calibration lane: is the spread honest — intervals and coverage. This is not a counterfactual about what would be worth knowing; it is the error we made, decomposed.

Evidence and next step

Worst positions by count: WR (7,954), RB (5,270), TE (5,089). READ THE DECOMPOSITION BEFORE PROPOSING ANYTHING. aar_attribution carries e_variance, e_regime, e_volume, e_share and e_conversion per forecast, so the question is not 'how do we get better' but 'why does THIS component dominate for THESE positions'. AND CHECK IT AGAINST THE RESIDUAL LEDGER, which answers a different question — what a variable would be worth if known perfectly, rather than what we actually got wrong. Where the two disagree is the interesting case: an error we keep making on a variable the ledger says is cheap means the cost is in our method, not in the unknown. A variance-labelled error is the honest residual of a stochastic process and may be correct; the question there is whether our INTERVALS cover it, which is calibration's layer, not whether the point estimate was wrong.

PW-48Probe an untested limit (cannot) in commit:0c3c6c18done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:0c3c6c18: "C-128: NYG@LAR kicks Mon 9/21 6:15 PM MT, so its first live close capture cannot be tested until then; last capture 4:41 PM MT today." PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-49Probe an untested limit (cannot) in commit:9800600ddone

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:9800600d: "cannot be fetched and must be imputed from snap counts, depth charts," PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-50Probe an untested limit (banned-floor) in commit:837568d5done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:837568d5: "senses actually present today: a measurement-procedure noise floor (refit" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-51Probe an untested limit (banned-floor) in commit:e4521890done

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:e4521890: "tdoe.py, actions.py, qarm.py, roleparam.py each compute a genuine noise floor" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-55Probe an untested limit (cost-estimate) in commit:0f5799bbdone

Owner: Claude · Lane: — · Priority: 1

Why this matters

bin/premise-scan.py found a sentence asserting something cannot be done, carrying no measurement. Sean, 2026-09-20: 'you assumed a limitation that doesn't exist.' An asserted limit is a claim, not a fact, and the cost of believing one is invisible -- the work is simply never attempted and nothing records that it was possible.

Evidence and next step

THE CLAIM, from commit:0f5799bb: "shipped in one commit (6b5b2d6a, same day, not multi-month), reading" PROBE IT. Either produce the measurement that makes it true -- a number, an interval, a sample size, the thing that turns an assertion into a finding -- or do the work it said could not be done. Record whichever happened. A limit that survives a probe is worth more than one that was never tested; a limit that does not was costing us the work.

PW-66BLOCKED_WITH_EVIDENCE: pay a lane for proving something impossibledone

Owner: Claude · Lane: operations · Priority: 1

Why this matters

'Every job has an authorized objective, impossibility detector, stop condition, and a BLOCKED_WITH_EVIDENCE terminal state. A worker earns credit for proving a task impossible; it never needs to manufacture success.' We have BLOCKED.md and a hold registry but no terminal state that PAYS for a proven impossibility. This is the designed answer to asserted-limits-are-claims, the failure mode with no trace. See §6.

PW-71Never grade on binary win/loss accuracy — magnitude is what paysdone

Owner: Claude · Lane: calibration · Priority: 1

Why this matters

Gemini's conflict resolution, settling two numbers in the dossier that look comparable and are not.

Evidence and next step

The dossier held a 1D CNN at '>74% accuracy' on 620 games against a JQAS study showing spreads capture '86% of the variability' on 5,000 matches. Different designs: the CNN evaluates BINARY classification, JQAS evaluates CONTINUOUS magnitude variance. Its ruling, which matches our own gate: 'You are staking CLV against continuous point spreads. Optimizing for binary accuracy guarantees building a model that perfectly mimics the moneyline and mathematically bleeds out to the vigorish.' Make it a rule the evaluation family enforces rather than a thing we remember: no promotion gate may be keyed to a win/loss hit rate.

PW-93Eight lanes starve the write lock: weather lost 8 hours of capturesdone

Owner: Claude · Lane: operations · Priority: 1

Why this matters

The watchdog caught it: 'weather capture — no work for 481 min (budget 480)'.

Evidence and next step

Two causes, one fixed. bin/weather-capture.py called duckdb.connect() directly, skipping the retry, the honest LockTimeout and D-106's 'a writer yields to the capture' — now routed through warehouse.connect and added to the capture exemption beside availability and availsnap, because a forecast at T-8h cannot be re-fetched at T-4h. THE SECOND CAUSE IS THE SCALING AND IT IS NOT FIXED. Eight build lanes plus their model fits contend for one DuckDB writer. Even through the connector the capture did not land on the retry after the fix. Reads do not contend; WRITES do, and more lanes means more of them. Options, none free: cap concurrent writers with a token so only N lanes may fit at once; give captures a reservation window the writers honour; or fewer lanes. Measure first — data/logs shows 14 build-lane lock failures today against 116 lock messages, so the contention is real but not yet the dominant cost. MEASURED 2026-09-23 01:35, from data/logs/job-misses.jsonl and long-holds.jsonl, since 20:00: 178 LOST FIRINGS, led by availability with 55 — the capture this project protects above everything, because ESPN serves only CURRENT status and a missed one is gone. Then gate 25, chain-report 18, ingest-rotowire 13, officials-capture 11, bodies 10. 60 LONG HOLDS, led by bodies with 18, then probe 10, rotowire-ids 5, news-sweep 5, stacked-arm 3. Q-01 recorded '14 retries over 157s all blocked by one PID holding the write lock 11+ min'. THE FAILURE IS NOT THAT WRITERS EXIST, IT IS THAT THEY ALL RETRY AT ONCE. Eight lanes and their fits hammer one file, each burning its whole budget, and the loser is whoever asked at the wrong moment — which is how an irreplaceable capture loses to a backfill that could have run any time. THE FIX IS A WRITE TOKEN: heavy writers queue for a token before attempting the DB lock, captures never wait for one, and a token whose owner died is reclaimed. That does not make writes faster; it makes the LOSER PREDICTABLE, and the loser must never be a capture. ATTEMPTED AND REVERTED 01:45. Editing warehouse.connect() live with 46 lanes running broke the module on an import indent and every lane would have failed on its next open. Restored from HEAD and verified. This belongs in a quiet window with the negative test written first: seed a long holder, prove a capture still lands. 2026-09-24 (CTO): priority 1 and blocks C-04 and N-10 -- the Monday-Wednesday early capture and the in-season refresh must not lose writes to lock starvation.

S-09S-09 — A learned penalty model: type, yards and automatic first down, by crew (C-07) and team disciplindone

Owner: Claude · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item S-09, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

A learned penalty model: type, yards and automatic first down, by crew (C-07) and team discipline rating. Defensive pass interference is real yardage that box scores record as zero. | Penalty yards per game by crew within 5%. FILES THE PLAN NAMES: research/playgrain/kernel/penalties.py MISSING: penalties.py No prior board item ever mentioned this.

S-12S-12 — Interception and fumble returns, defensive TDs, recovery. This replaces the additive 2.18 pts/tedone

Owner: Claude · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item S-12, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Interception and fumble returns, defensive TDs, recovery. This replaces the additive 2.18 pts/team-game constant in `drives.py:865`. | Non-offensive points emerge within 0.2. FILES THE PLAN NAMES: research/playgrain/kernel/turnover_returns.py MISSING: turnover_returns.py No prior board item ever mentioned this.

S-13S-13 — Rule-exact overtime by era. | 61% of overtime games finish on 3, as in history.done

Owner: Claude · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item S-13, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Rule-exact overtime by era. | 61% of overtime games finish on 3, as in history. FILES THE PLAN NAMES: research/playgrain/kernel/overtime.py MISSING: overtime.py No prior board item ever mentioned this.

S-14S-14 — The rollout, vectorised across trajectories and the slate, with **player attribution**: every pldone

Owner: Codex · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item S-14, lane kernel, drift, decision, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The rollout, vectorised across trajectories and the slate, with **player attribution**: every play records passer, carrier and target, so player stat lines accumulate per trajectory. It draws lineup, rating posterior, form, weather and crew per trajectory. | One ledger: on every path, QB passing equals the sum of receivers, and team points equal the sum of scores. Exactly. FILES THE PLAN NAMES: research/playgrain/sim2.py MISSING: sim2.py No prior board item ever mentioned this. CURRENT DELIVERY WORK: REOPENED 2026-09-24 delivery audit: known offensive ledger, fresh starter selection and QB ability are now tested; full-game overtime, passing fumbles, return/two-point participant attribution and K/DST are unfinished. Overtime possessions and final scoring-entity conservation are now measured (research/playgrain/s14_overtime_2024.json). Passing-fumble mechanics now measured with opt-in per-team hazards; both held-out seasons and the full rollout overpredict losses, so calibration/promotion remains unresolved. Next bounded unit: explicit return/two-point participant attribution. K/DST and third-postseason-period kickoff election remain unsupported. See research/playgrain/s14_pass_fumbles_unit.json. See docs/institute/PROSPECTIVE-SHADOW-2026-09-24.md. Do not certify partial regulation fantasy as full-game predictions. Successful two-point participant attribution is now opt-in and mechanically measured across all 272 held-out 2024 games. Participant-selection accuracy is unvalidated. Next bounded unit: return touchdown participant identity. S-14 remains incomplete; see research/playgrain/s14_two_point_unit.json. Kickoff-return touchdown identity is now opt-in and mechanically measured across 272 2024 games. Prior-history returner support is only 72.4%/79.1% in 2024/2025, so the selector is not promoted. Next bounded unit: separate ordinary punt-return TDs from blocked-punt recoveries before assigning participants. Existing special-teams live rates include 2026 data; rollout conservation is a mechanics check, not fully held-out forecast accuracy. S-14 remains incomplete; see research/playgrain/s14_kick_returners_unit.json. Punt return TDs and blocked-punt recovery TDs are now distinct, with mutually exclusive block outcomes and penalty cancellation measured. Participant identity remains unresolved. Next bounded unit: ordinary punt-returner identity, excluding muff/kicking-team touchdown rows; blocked recoverers need their own policy. See research/playgrain/s14_punt_events_unit.json. S-14 remains incomplete. Ordinary punt-return TD identity is now opt-in and mechanically verified on 35 2024 games spanning all 18 weeks. Prior-history identity support is 85.7%/85.6% in 2024/2025; stale histories can nominate a player to both opponents, which the public rollout now refuses. One such matchup was excluded from mechanics. No promotion. Next bounded unit: blocked-punt recovery participant identity; returner availability reconciliation is also unresolved. See research/playgrain/s14_punt_returners_unit.json. S-14 stays incomplete. Blocked-punt touchdown scorer identity is now opt-in and mechanically verified on 36 games across all 18 weeks of 2024. Prior same-team scorer history supports 0/3 and 0/2 touchdowns in 2024/2025, so no selector is promoted. Empty histories explicitly retain team credit. Next bounded unit: reconcile returner candidates with clocked team availability; K/DST and other full-game gaps remain. See research/playgrain/s14_blocked_punt_scorers_unit.json. S-14 stays incomplete. Returner roster eligibility is now opt-in using observed state/event clocks. No captured evidence exists for 2024/2025, so held-out identity coverage is zero and the policy is not promoted. All 36 sampled 2024 games retain exact ledgers with unknown identities at team level. Next bounded unit: kicker identity and field-goal/PAT attribution. Injury-aware availability and other full-game gaps remain. See research/playgrain/s14_returner_availability_unit.json. S-14 stays incomplete. Kicker identity and made-PAT attribution are now opt-in and mechanically measured across 36 games spanning all 18 weeks of 2024. Prior-game identity accuracy is 92.70%/92.98% in 2024/2025, but availability is unvalidated; no promotion. Next bounded unit: typed try outcomes for PAT attempts/misses. Distance-specific kicking statistics, DST and other full-game gaps remain. See research/playgrain/s14_kickers_unit.json. S-14 stays incomplete. PRIORITY 0 -> 1 2026-09-24 (CTO, build audit /tmp/claude-501/build-audit-2026-09-24.md): what remains is player attribution for fantasy and props; the game price already includes overtime (sim.py:791). Continue after S-06.

T-01T-01 — The role of every player on every play: passer, carrier, target, route runner, pass blocker, rundone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 1

Why this matters

Plan item T-01, lane ratings, lineups, policies, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The role of every player on every play: passer, carrier, target, route runner, pass blocker, run blocker, rusher, coverage, box. From participation and pbp ids. | Coverage ≥ 99% of 2016–2025 snaps. FILES THE PLAN NAMES: research/playgrain/talent/player_roles.py MISSING: player_roles.py No prior board item ever mentioned this.

T-04T-04 — Priors for players with little history: draft pick (the only B-hypothesis that ever paid, B13/B4done

Owner: Claude · Lane: ratings, lineups, policies · Priority: 1

Why this matters

Plan item T-04, lane ratings, lineups, policies, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Priors for players with little history: draft pick (the only B-hypothesis that ever paid, B13/B41: +0.0475, +0.0595), combine, college production (C-17), Madden ratings (`data/madden.csv`), position age curves. | Rookie first-four-games error beats the league-mean prior. FILES THE PLAN NAMES: research/playgrain/talent/priors.py, talent/age_curves.py MISSING: priors.py, age_curves.py No prior board item ever mentioned this.

T-09T-09 — Shrinkage audit DONE: five families, each with a held-out comparisondone

Owner: Claude (closer) · Lane: unassigned · Priority: 1

Why this matters

Gemini 1/10: small-sample skill estimates inflate variance; the QB rating shrinks, other families may not.

Evidence and next step

MEASURED PARTIAL 2026-09-25: frozen game-score kernel reads four team EPA latents + QB + seen; E-01 historical runner supplied QB=0. In as-of next-game EPA proxy tests on 2024/25, removing the explicit team K=4 prior did not improve any of four components in either season. Removing QB K=4 improved the n<4-history QB bucket in both years (2024 n=45, MSE 0.23026→0.22043; 2025 n=55, 0.22832→0.20960). These are unit-signal diagnostics on already-used years, not game-score or price gains; no parameter or release was promoted. T-09 remains OPEN pending QB-inclusive full-game rerun and matched kernel shrinkage comparison. Shrinkage audit (Gemini 1/10, 2026-09-25): every skill vector the kernel reads (QB, rusher, target, blocker, rusher-D, coverage, kicker) is checked for small-sample treatment the way the QB rating already has it (shrink_games, half-life). Report per family: effective n, prior, weight n/(n+K), and the held-out gain of shrinking vs not; add empirical-Bayes shrinkage where a family has none. ACCEPTANCE: Every family listed with its shrinkage and a held-out comparison; no family feeds raw small-sample means to the kernel. UNBLOCKED 2026-09-25 07:40 MT (Claude): Codex finished its overnight T-09 work (06:24). Findings on record: team-prior over-shrinkage REFUTED (fixed grid, 2024+2025); novice-QB over-shrinkage open. Close this item: write up the novice-QB result with its interval and mark done; any fix is a kernel candidate, not this item.

V-06V-06 — No deploy may break a rendered invariantdone

Owner: Codex · Lane: construction · Priority: 1

Why this matters

bin/verify-live.py checks that markers are PRESENT after a deploy, and presence is not correctness — it was green throughout the week-1 defect. The invariants are V-01's rules, run against the bytes the browser receives, and a deploy that breaks one is rolled back. EXAM: seed a broken render and watch the deploy refuse. A check never seen red is not a check.

X-03X-03 — Density-ratio classifier (real states against simulated); reweight the transition fits to the sidone

Owner: Claude · Lane: kernel, drift, decision · Priority: 1

Why this matters

Plan item X-03, lane kernel, drift, decision, due Fri 25 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Density-ratio classifier (real states against simulated); reweight the transition fits to the simulator's state distribution. | Play-level log-loss on real states not worse by more than the tax. FILES THE PLAN NAMES: research/playgrain/drift/dagger_refit.py MISSING: dagger_refit.py No prior board item ever mentioned this.

B-02B-02 — Action model v2, replacing `train_action.py`: P(action \| state, personnel, unit ratings, QB, wedone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 2

Why this matters

Plan item B-02, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Action model v2, replacing `train_action.py`: P(action \| state, personnel, unit ratings, QB, weather, play-caller effect). Multiclass LightGBM with as-of hierarchical play-caller encodings. Includes pass rate over expected, tempo, shotgun, play-action, screen, RPO and motion rates. Play-caller intent is modelled separately from QB execution: a checkdown is not a conservative call. FILES THE PLAN NAMES: research/playgrain/policy/policy_offense.py MISSING: policy_offense.py No prior board item ever mentioned this.

C-08C-08 — Two to four named beat writers per team through RSS, Bluesky and team sites, via `playerweek/brodone

Owner: Claude · Lane: sensors · Priority: 2

Why this matters

Plan item C-08, lane sensors, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Two to four named beat writers per team through RSS, Bluesky and team sites, via `playerweek/browser.py` (isolated headless browser, never your Chrome). Received-time stamped. FILES THE PLAN NAMES: research/playgrain/sensors/beatwire_capture.py MISSING: beatwire_capture.py PRIOR EVIDENCE on the old board: C-08

PW-199H-455 — The v0.1 game total is over-dispersed: actual on predicted regresses with slope 0.61 for the total and 0.76 for team points (2018-21; total 0.75/0.57/0.58/0.41 by season), the totadone

Owner: — · Lane: — · Priority: 2

Why this matters

Tested-positive research card H-455 (research/scientist/HYPOTHESES.md, status CLAIMED) has no implementation unit on the board yet. [card node=G.total track=A rollup_wpp=0.02 p_real=0.5 cost=1 residual=v01-total-over-dispersion] **The v0.1 game total is over-dispersed: actual on predicted regresses with slope 0.61 for the total and 0.76 for team points (2018-21; total 0.75/0.57/0.58/0.41 by season), the total-side twin of H-413's margin slope 0.82.** MECHANISM: teamstate/gametotal learn each team's scoring level from final scores with no shrinkage stage, so a level carries the noise of a few high- or low-scoring games and the predicted total spreads wider than the outcomes it forecasts. PRIOR ART: decision 2026-10-07T0159Z

Evidence and next step

Filed by bin/feed-predictive-queue.py from HYPOTHESES.md CLAIMED card. Acceptance is a GRADED RESULT on the shared point-in-time harness (winner % vs Vegas with n and interval, or the relevant metric-tree node) -- never code merged.

PW-201H-468 — Bridge: Draft position helps forecast players with little NFL history (B13, B41)done

Owner: — · Lane: — · Priority: 2

Why this matters

Tested-positive research card H-468 (research/scientist/HYPOTHESES.md, status CLAIMED) has no implementation unit on the board yet. [card node=G.winner track=B rollup_wpp=0.03 p_real=0.3 cost=2 residual=bridge:2026-09-11-draft-capital-thin-history-b13] **Bridge: Draft position helps forecast players with little NFL history (B13, B41)** MECHANISM: research/experiments/2026-09-11-draft-capital-thin-history-b13.md (Track A, Promoted) must enter the World model as a mechanism, not a points fudge. Name which: who takes the snaps/role on the field, a style parameter the simulator already carries, or a decision policy (play-calling, fourth-down, target/carry allocation). State the mechanism and the kernel component it changes. EV

Evidence and next step

Filed by bin/feed-predictive-queue.py from HYPOTHESES.md CLAIMED card. Acceptance is a GRADED RESULT on the shared point-in-time harness (winner % vs Vegas with n and interval, or the relevant metric-tree node) -- never code merged.

T-03T-03 — Directly measured component skills with shrinkage, from NGS, PFR advanced, FTN and pbp (the compdone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 2

Why this matters

Plan item T-03, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Directly measured component skills with shrinkage, from NGS, PFR advanced, FTN and pbp (the component list follows this table). | Each component must earn its place through L-10. FILES THE PLAN NAMES: research/playgrain/talent/skill_components.py MISSING: skill_components.py No prior board item ever mentioned this.

T-06T-06 — The lineup → unit function: a learned map from the on-field players' vectors to the unit quantitdone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 2

Why this matters

Plan item T-06, lane ratings, lineups, policies, due Tue 29 Sep. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The lineup → unit function: a learned map from the on-field players' vectors to the unit quantities the kernel consumes (pass protection, rush, coverage, run block, run fit, QB, skill group). | Natural-experiment test: across every historical starter absence, predicted unit change against realised. Slope 0.8–1.2. FILES THE PLAN NAMES: research/playgrain/talent/unit_compose.py MISSING: unit_compose.py No prior board item ever mentioned this.

B-07B-07 — A play-caller effect has to survive a change of team and quarterback, or it is a team effect in done

Owner: Claude · Lane: ratings, lineups, policies · Priority: 3

Why this matters

Plan item B-07, lane ratings, lineups, policies, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

A play-caller effect has to survive a change of team and quarterback, or it is a team effect in disguise. FILES THE PLAN NAMES: research/playgrain/policy/policy_transport_test.py MISSING: policy_transport_test.py No prior board item ever mentioned this.

C-15C-15 — Every public Big Data Bowl tracking release (10 Hz, all 22 players) into `data/tracking/*.parquedone

Owner: Claude · Lane: sensors · Priority: 3

Why this matters

Plan item C-15, lane sensors, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Every public Big Data Bowl tracking release (10 Hz, all 22 players) into `data/tracking/*.parquet`. Used to train T-07, never joined to live weeks. FILES THE PLAN NAMES: research/playgrain/sensors/bdb_ingest.py MISSING: bdb_ingest.py PRIOR EVIDENCE on the old board: C-15, C-77

E-10E-10 — Re-run what the old path measured wrongly (the queue follows this table).done

Owner: Claude · Lane: truth and the fantasy path · Priority: 3

Why this matters

Plan item E-10, lane truth and the fantasy path, due Fri 2 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Re-run what the old path measured wrongly (the queue follows this table). FILES THE PLAN NAMES: research/playgrain/evalx/retest_queue.py MISSING: retest_queue.py No prior board item ever mentioned this.

PW-56Grade the STAT LINE, not fantasy pointsdone

Owner: Claude · Lane: forecast · Priority: 3

Why this matters

Sean, 2026-09-22: "actuals are the stats, fantasy points is math."

Evidence and next step

statline.score() is a dot product of a 33-field StatLine with ONE league's scoring vector — it calls itself 'the only such function'. The model computes the whole line; every grading surface discards it and stores the projection. aar_forecasts HAS a `quantity` column built for this and every writer hardcodes 'fantasy_points'. Two different lines with the same total grade identically, so a miss cannot be attributed to which stat was wrong; and a prop is priced on a stat, so an edge there is invisible in points MAE. player_week already carries every actual column (targets, receptions, receiving_yards, carries, rushing_yards, TDs, first downs, fumbles). bin/aar-ladder.py is the start.

PW-63L-07 promotes its own proposals; promotion is not its authoritydone

Owner: Claude · Lane: decision · Priority: 3

Why this matters

'A lower model may falsify or quarantine a frontier result. It may not promote the result. Promotion comes from instruments and an independent reviewer.' The loop promotes unattended and has done so successfully, which is why it never looked wrong. Falsification authority and promotion authority are different powers. See §3.

L-08L-08 — The second loop. The ledger's top unexplained cluster becomes a brief naming two worlds the machdone

Owner: Claude · Lane: sensors · Priority: 4

Why this matters

Plan item L-08, lane sensors, due Tue 6 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

The second loop. The ledger's top unexplained cluster becomes a brief naming two worlds the machine cannot tell apart and the smallest observation that would separate them. Deep Research is driven through `browser.py`; its answer passes the airlock (schema, citations must resolve) or is dropped. `source_probe.py` fetches a sample and measures latency and rights. An admitted source becomes a capture config. FILES THE PLAN NAMES: research/playgrain/loop/sensor_loop.py, loop/intake_airlock.py, loop/source_probe.py MISSING: sensor_loop.py, intake_airlock.py, source_probe.py No prior board item ever mentioned this.

T-07T-07 — From the tracking releases: speed, acceleration, separation at throw, get-off, closing speed. Thdone

Owner: Claude · Lane: ratings, lineups, policies · Priority: 4

Why this matters

Plan item T-07, lane ratings, lineups, policies, due Tue 6 Oct. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

From the tracking releases: speed, acceleration, separation at throw, get-off, closing speed. Then a teacher-student model predicts those from public stats, so every player in every season gets an estimate. | Student R² reported per skill; enters only via L-10. FILES THE PLAN NAMES: research/playgrain/talent/tracking_latents.py MISSING: tracking_latents.py No prior board item ever mentioned this.

F-07F-07 — Turn on `K_PER_POSITION` and `POSR_DRIFT`; replace the covariate-free shrinkage with conditionaldone

Owner: Claude · Lane: unassigned · Priority: 9

Why this matters

Plan item F-07, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Turn on `K_PER_POSITION` and `POSR_DRIFT`; replace the covariate-free shrinkage with conditional models on the same features. | Scored unconditionally FILES THE PLAN NAMES: none all present No prior board item ever mentioned this.

O-04O-04 — Every new job beats a heartbeat, and silence raises an alarm on the board.done

Owner: Claude · Lane: unassigned · Priority: 9

Why this matters

Plan item O-04, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

Every new job beats a heartbeat, and silence raises an alarm on the board. FILES THE PLAN NAMES: operate/watch.py all present No prior board item ever mentioned this.

P-11P-11 — `asof()` covers every new table. | 7/7 → n/n.done

Owner: Claude · Lane: unassigned · Priority: 9

Why this matters

Plan item P-11, lane unassigned, due unscheduled. This board IS the plan now: 119 items, their files, their order.

Evidence and next step

`asof()` covers every new table. | 7/7 → n/n. FILES THE PLAN NAMES: warehouse.py, test_warehouse.py all present No prior board item ever mentioned this.

PW-193Wire the Desk's GPU transcriber into the podcast pipelinedone

Owner: Claude · Lane: — · Priority: —

Why this matters

The Desk (192.168.0.16, GTX 1070) transcribes at ~19x realtime, faster than all of Nimo combined (~11x), and is proven working (f409d5fc7). It sits idle because nothing feeds it. Transcripts feed the B/Q thought-leader lanes (QUALITATIVE-SOURCES.md). Sean gave full control of the machine on 2026-10-06 so it works as a free server.

Evidence and next step

BUILT AND PROVEN LIVE 2026-10-06. bin/desk-feed.py (new, sibling to bin/nimo-feed.py) claims episodes from the shared data/transcripts/priority-order.json in existing tier/priority order, skipping anything already transcribed or live-claimed by Nimo (nimo-claims.json) or the Mac (mac-working.json); downloads the audio itself and stages it straight into the Desk's ready/ + archive/<key>.audio over SFTP; records its own claims in desk-claims.json; and collects out/ and failed/ receipts through the SAME validate()/atomic-write-if-absent path bin/nimo-feed.py uses (imported directly, not re-implemented, so the two can never drift). bin/nimo-feed.py's own staging loop and research/podcast_transcribe.py's queue (desk_claimed(), mirroring nimo_claimed()) both skip a key the Desk already holds, so every episode is claimed by at most one host; Nimo's running workers were never stopped or reconfigured. TARGET_HOURS is the same constant nimo-feed.py uses, scaled by the Desk's own (much higher) measured rate, which is what gives it the biggest share of the queue without separate split logic -- a single live pass claimed all 382 current-season (T0) episodes already held by Nimo were correctly skipped, and the Desk took over 1,000 backfill-lane episodes instead. WORKER COUNT, measured live (ops/desk/README.md 'GPU headroom and worker count'): 1 GPU worker alone ran ~18-19x realtime (matches the original 19.3x isolated benchmark); 2 GPU workers at --threads 2 each plus 1 CPU worker at --threads 4 (8 of 8 logical i7-7700 threads, no oversubscription) measured ~20.6x combined on the GPU side -- a real but modest gain. Adding a second CPU worker on top (12 requested threads on 8) was measured WORSE: GPU throughput fell to ~15.8x combined because the CPU-side VAD/decode glue the GPU workers need got starved; reverted. Deployed: gpu0, gpu1 (2 threads each), w0 (4 threads) -- install-workers.ps1 now takes -GpuThreads separately from -CpuThreads for this. Also measured and documented: reinstalling scheduled tasks does not reliably stop an already-running worker process; the fix (kill the python/cmd processes directly, then reinstall) is recorded in ops/desk/README.md. data/desk-status.json (bin/desk-feed.py, every 5 min) mirrors nimo-status.json's shape. bin/lane-watchdog.py's new desk_rows() adds 'desk reachability' and 'desk live transcriber' rows into the same record_incidents() path Nimo's rows use (routed to the headless fixer); both checks were proven to go dead against a planted bad status file, then clean again, before being trusted. buildsite/pipeline_progress.py's _desk() and buildsite/desks.py's pipeline_panel() put a Desk row next to Nimo's on the build site's Data page. bin/desk-maintenance.py (new, launchd Tuesdays 03:00 MT) installs Windows security updates, reboots and verifies both tasks come back; it fails closed on an unreadable warehouse or any game within 6h (never runs in a game window), and self-gates to a monthly cadence despite the weekly trigger. Proven live: a real, previously-untranscribed episode was claimed, downloaded, transcribed on the Desk, and collected into data/transcripts/raw/ with zero duplicates in the same session; bin/nimo-feed.py ran a clean pass afterward with the new skip clause and did not disturb Nimo's held work.

Research intake

7 registered · 1 measured, 6 queued · source Sep 23 · 9:29 PM MT · Open World Model Lab for records.