17 Nov 2026 · 10:44 AM MTUpdated…
World model lab

One world. Every outcome.

Model the game one snap at a time; learn which missing mechanisms improve the predictions.

Forecast accuracy ↗

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.

Where better information could help

Perfect realized values are inserted one at a time. This ranks research opportunities; it is not a measured production gain.

VariableOracle improvementMAE with oracleSensor
his snap share0.70253.6496role forecasting from depth, usage and the injury ladder
his targets0.43963.9125the target policy: who gets the ball, given who is on the field
did he play at all0.41393.9382the availability hurdle (C-53) -- P(active) before kickoff
his carries0.19944.1527the carry policy and the backfield committee split
team pass attempts-0.06394.4159team volume: plays and pass rate given script and opponent
team rush attempts-0.08204.4341team volume, rushing half

Residual ledger · Nov 17 · 10:21 AM MT

Gemini research: decisions and experiments

7 registered · 1 measured, 6 queued · source Sep 23 · 9:29 PM MT

QueuedH-264

Weather × surface × type of play

The existing weather retest has intervals spanning zero. Test conditional effects and forecast vintages before discarding the mechanism or adding fixed penalties.

Test, data and promotion criteria

Use stored forecast runs available at each lock; compare a pooled baseline with wind/rain/temperature × roof/surface × pass depth or kick distance. Train through 2022, tune on 2023, reserve 2024–25 where forecast archives exist. Cluster uncertainty by game; control the family of interactions and test shuffled weather. If no holdout remains untouched, require prospective confirmation.

Data

Existing Open-Meteo lead ladder and stadium metadata. Exact model runs need publication lag and capture time. No historical realized-weather substitution.

Plan: C-06 · E-10 · Lane: s-observation · Related: H-013

QueuedH-265

Workload, recovery and body-clock interactions

Travel features already exist. The new question is whether prior snaps, short rest and local kickoff jointly change late-game execution, rather than applying a universal eastbound penalty.

Test, data and promotion criteria

Pre-register a small interaction family using lagged snaps, days rest, timezone shift and local kickoff. Evaluate by held-out season and game with team/player shrinkage; compare against rest-only and travel-only baselines plus shuffled travel. Never derive a neurological diagnosis or millisecond penalty from collision counts.

Data

Existing schedules, participation and play-by-play; exposure features must precede the prediction lock.

Plan: T-03 · B-02 · E-10 · Lane: b-judgement · Related: H-028

QueuedH-266

How coaches adapt to an opponent

Sequence memory is relevant to a snap-by-snap world. A 3B network or game-tree search is not justified before smaller sequence features beat the current policy.

Test, data and promotion criteria

Compare current policy against lagged play-family and opponent-response summaries, with game-separated rolling origins. Hidden calls stay hidden from the opposing policy. Score action log loss and rollout score/player distributions with common seeds. Shuffled sequence order is a negative control. No promotion from action accuracy alone.

Data

Public play-by-play supports coarse sequences. Full 22-player raw tracking is a separate, limited-access input.

Plan: B-02 · B-05 · L-10 · Lane: b-judgement · Related: new intake

QueuedH-267

Player matchups with honest observability

Available charting can support partially pooled line and coverage matchups. An exact physics graph requires measurements that public summary statistics do not contain.

Test, data and promotion criteria

Inventory existing charting and ID crosswalks first. Compare player-pair effects with team/unit priors on held-out players and seasons; require gain after shrinkage and a shuffled-matchup control. Freeze availability and role at the prediction horizon.

Data

Existing licensed/owned charting first; NFL Big Data Bowl for bounded tracking research, not an assumed live all-game feed.

Plan: T-01 · B-05 · S-19 · Lane: s-observation · Related: new intake

QueuedH-268

Officiating as an observable mechanism

Penalties and reviews are observable. Inches of referee spotting bias are not established by the report, and line-to-gain measurement technology has changed.

Test, data and promotion criteria

Test partially pooled crew effects on penalties and reviews, conditioned on teams, venue and era. Forward-season evaluation and crew-label placebo required. No continuous spotting offset without a validated location label.

Data

Play-by-play penalties/reviews and crew assignments with source timestamps. Crew feed coverage must be measured first.

Plan: S-09 · E-10 · Lane: s-observation · Related: new intake

MeasuredE-02

Score concentration must earn its confidence

Repeated simulation agreement is not an observed prediction success rate. The archived draws now have a reproducible exact-score diagnostic.

Test, data and promotion criteria

Track modal score probability against modal hit frequency and joint-score Brier score. Repeating draws must not improve a score; confidently wrong collapsed draws must score worse. Compare candidates on untouched or prospective games before promotion.

Data

4,897 archived rolling-origin games, 150 draws each; source hashes and historical-rule caveats in the evidence file.

Plan: E-02 · Lane: calibration · Related: new intake

QueuedH-269

Profile the simulator before changing languages

The existing attribution benchmark misses its target under heavy machine contention. The report supplies no measured evidence for a mandatory Rust rewrite or its memory budget.

Test, data and promotion criteria

Repeat O-05 under recorded contention and an idle comparison. Attribute setup, booster prediction, per-snap Python and ledger costs. Optimize the dominant segment, preserve paired-seed distributions and ledger conservation, then rerun the 16-game × 10,000-draw workload.

Data

Existing bench.json, timing profiler and actual device memory. MLX is an option if a measured neural workload warrants it.

Plan: O-05 · S-19 · Lane: construction · Related: new intake

Architecture decisions
Build decision

Keep one snap-by-snap world. Every game, prop and fantasy output comes from the same trajectories.

Build decision

Keep market prices outside the explanatory core. Closing prices cannot be features for earlier prediction locks.

Build decision

Preserve the original forecast, code/data versions and actual capture time at every lock. Rebuilt past games are retrospective.

Build decision

No fixed injury, circadian, glove, turf or aerodynamic coefficient is promoted from this report without measured inputs and held-out evidence.

Build decision

No claimed 81–99% exact-score convergence or staking recommendation in the report is treated as an observed result.

Build decision

Full tracking feeds, ball microphysics and neural latency remain data-acquisition questions, not implemented sensors.

Primary references

The research report is an input to review, not evidence of the performance it claims.

Throughput is a measured build problem

The saved O-05 benchmark reports OVERBUDGET under heavy load. O-05 / S-19 will profile and optimize the existing simulator before deciding whether compiled or GPU execution earns its cost.

Season simulation research
Model lineage

rating margin model, not the snap kernel (S-18 NOT-DONE 1)

TeamExpected winsPlayoffsDivisionTitle
JAX12.7899.3%97.5%16.5%
BUF13.5599.7%97.3%16.0%
SF12.9195.9%80.0%11.0%
DET11.9388.2%39.1%7.4%
CIN12.4597.8%78.8%7.2%
MIN11.9891.3%39.8%6.4%
CHI11.2983.0%19.8%5.7%
LA10.1359.3%9.6%5.7%
KC11.6895.1%46.9%5.6%
DEN12.1194.9%52.2%5.2%
SEA10.0847.3%10.5%3.4%
PIT10.3976.0%12.2%1.7%
TB7.5647.4%47.2%1.7%
NYG8.9359.4%58.2%1.4%
GB9.3131.3%1.2%0.9%
HOU7.6617.7%2.3%0.8%
NE8.5445.1%2.7%0.7%
CAR7.4734.4%33.9%0.7%
DAL7.2519.9%19.3%0.6%
WAS7.1321.0%20.4%0.4%
CLE9.2256.7%8.6%0.4%
NO6.1511.4%11.2%0.2%
ATL5.717.8%7.7%0.1%
PHI5.062.1%2.1%0.1%
BAL6.687.0%0.4%0.1%
LV6.838.2%0.8%0.0%
IND5.741.9%0.1%0.0%
ARI5.840.1%0.0%0.0%
TEN3.520.1%0.0%0.0%
NYJ4.210.2%0.0%0.0%
MIA3.720.1%0.0%0.0%
LAC4.180.2%0.0%0.0%

Season simulation · Nov 16 · 10:36 PM MT

Build commitments

Source: executable build plan. A completed item is not proof of betting edge.

ItemStateWork
S-16doneS-16 — Trajectories are content-hashed per (game, lock, release); issuance is frozen with the existing
S-11doneS-11 — Kickoffs (with the 2024 rule era), punts, returns, blocks, and field goals by kicker rating × di
T-02doneT-02 — Regularised adjusted plus-minus at the play grain: a sparse design matrix (players on field × ro
X-02doneX-02 — Add `form_off`, `form_def` (leave-one-play-out EPA in this game) to `LATENT`; fit the distributi
X-04doneX-04 — Simulation-based calibration of a small vector of logit offsets against drive and game moments f
B-03doneB-03 — Personnel grouping (11/12/21…) given state and roster. This selects the on-field set from A-05.
T-05doneT-05 — Weekly state-space update per player-skill: process noise by age and position, a shock term on r
X-05doneX-05 — Key numbers: the mass at 3, 6, 7, 10, 14 and 17 on margin, and at 37, 41, 44, 47 and 51 on total
X-06doneX-06 — The column-by-column real-against-simulated feature diff that found the 3.2-sd `games_seen` shif
B-05doneB-05 — Box count, rushers, man/zone, shell, given state, offensive personnel and DC. These are the W-00
B-06doneB-06 — Who gets the ball. Run: carrier among on-field backs and QB (designed or scramble). Pass: a mult
S-17doneS-17 — Any team-season against any team-season, retired players included. The north star's first demand
S-18doneS-18 — Whole seasons and playoffs. Ratings evolve between simulated weeks and injuries accrue through A
S-19doneS-19 — Compile the boosters (treelite or lleaves, or Numba lookup tables) if S-14 misses its budget. |
T-08doneT-08 — Every rating family is scored on what it adds to the kernel and logged to the ledger. | —
S-15doneS-15 — The compiler: margin, total, moneyline, team totals, alt lines, any prop, fantasy points under a
S-14doneS-14 — The rollout, vectorised across trajectories and the slate, with **player attribution**: every pl
T-09doneT-09 — Shrinkage audit DONE: five families, each with a held-out comparison
B-02doneB-02 — Action model v2, replacing `train_action.py`: P(action \| state, personnel, unit ratings, QB, we
X-01doneX-01 — **First thing built.** k-step rollouts from real states (k = 1, 2, 4, 8, 16, to end of drive); d
S-08doneS-08 — Early stopping on a validation season, not an inference cap at 60 rounds. 115 integer yard bins.
T-01doneT-01 — The role of every player on every play: passer, carrier, target, route runner, pass blocker, run
S-12doneS-12 — Interception and fumble returns, defensive TDs, recovery. This replaces the additive 2.18 pts/te
S-09doneS-09 — A learned penalty model: type, yards and automatic first down, by crew (C-07) and team disciplin
X-03doneX-03 — Density-ratio classifier (real states against simulated); reweight the transition fits to the si
T-03doneT-03 — Directly measured component skills with shrinkage, from NGS, PFR advanced, FTN and pbp (the comp
T-06doneT-06 — The lineup → unit function: a learned map from the on-field players' vectors to the unit quantit
S-13doneS-13 — Rule-exact overtime by era. | 61% of overtime games finish on 3, as in history.
S-10doneS-10 — Semi-Markov durations: the elapsed-time *distribution* per play type, state, tempo and coach, in
B-04doneB-04 — Go/punt/kick by coach. The late-and-close policy is its own model: kneel-downs, playing for the
B-07doneB-07 — A play-caller effect has to survive a change of team and quarterback, or it is a team effect in
T-07doneT-07 — From the tracking releases: speed, acceleration, separation at throw, get-off, closing speed. Th
T-04doneT-04 — Priors for players with little history: draft pick (the only B-hypothesis that ever paid, B13/B4
B-01doneB-01 — Head coach, play-caller, OC and DC per team-week, as-of. `coordinators.csv` exists; fix `games_r
S-06doneS-06 — Calibration measured: held-out reliability FAILS

Desk roster

7 staffed, 1 partial, 0 struck. Every job the operating model charters for this desk (docs/WHAT-I-WAS-MISSING.md #4), against what actually runs in this repo -- not what the tab above implies.

JobStatusEvidence
Chief World-Model Architect
Defines how observation, latent state, policies, plays, games, seasons, and decisions compose.
staffeddocs/PLAN-OF-RECORD.md names the composition; playerweek/sim.py and sim_world.py implement the rollout it describes.
Bayesian Forecasting Scientist
Represents uncertainty, partial pooling, priors, and posterior updates without false precision.
staffedplayerweek/intervals.py (shipped_band, partial-pooling floor) and research/playgrain/talent/rapm_fit.py (ridge shrinkage) represent uncertainty without full precision.
Causal Inference Economist
Separates predictive proxies from mechanisms and designs credible interventions.
staffedresearch/measure_synthetic_control.py and the T-06 natural-experiment slope (coach-change as a natural experiment) separate proxy from mechanism.
Time-Series & Regime Scientist
Finds temporal structure, change points, decay, and schedule-dependent functions from the bottom up.
staffedplayerweek/regime.py, playerweek/desigdecay.py, and research/measure_alpha_decay.py find change points and decay from the data.
Multi-Agent Simulation Scientist
Builds opposing team policies that learn and adapt inside legal football.
staffedresearch/playgrain/policy/policy_offense.py (B-02, beats v1 3/3 holdout seasons) and research/build_defense_policy.py give both sides of the ball a learned, adaptive policy inside playerweek/sim.py's rollout.
Continuous Deep Research Investigator
Keeps a cross-domain research queue moving without waiting for Sean to supply ideas.
staffedresearch/scientist/ runs a self-directed hypothesis queue (HYPOTHESES.md, RANKED.md, prereg/) without waiting for Sean to supply ideas.
Physics & Spatial Dynamics Scientist
Models movement, contact, leverage, ball flight, surface, and energy constraints.
partialplayerweek/ngs.py ingests Next Gen Stats tracking measures (separation, etc., 2016-2025) and research/playgrain/talent/tracking_latents.py (T-07) fits tracking-derived skill latents -- but on ONE 2017 game (316,025 frames; 7 of 8 public releases 403), and no movement, contact, or ball-flight model exists. [independent review 2026-09-23: was struck]
Calibration & Uncertainty Scientist
Tests whether probabilities, intervals, and joint simulations mean what they claim.
staffedplayerweek/calibrate.py, playerweek/calibration.py, and research/measure_calibration*.py test whether intervals and joint simulations mean what they claim.