Recent findings, experiments and independent reviews from the research lanes. Each record shows its finding and next action. Current live improvement remains unverified by this list.
13Lane findingsResearch and experiment records, last 7 days
0Independent reviewsReferee records, last 7 days
8Adopted in registerHistorical decisions; not necessarily served
?Current live linkage0 listed in register verified 2026-09-25; current release proof unavailable
Record timing and evidence limits
“Today” uses Mountain calendar dates. Times come from record IDs in UTC and are shown in Mountain time; they are not verified completion times. A record identifies its lane, not which paired scheduled job alias ran it. Unknown owner lanes stay unassigned. 13 total records = 13 lane findings + 0 independent reviews. A finding can mix positive and negative conclusions, so its authored verdict appears in full rather than a guessed pass/fail label. Adoption and a live shipped gain require separate release proof. 0 unreadable eligible records were omitted.
The served kernel still draws defensive nullifying fouls at a flat rate by down (3rd-down ratio 0.986 against a real band of 1.78-2.16); this is a rules and fidelity finding about the kernel, not a forecast result, and the P04 audit item stays unverified only because its own checker plant landed 0.04 above the band.
Rejected as new information: points_l3_diff on play EPA is absorbed by the as-of team latents (held-out slope 0.041 to 0.009, CI spans 0). The scan baseline lacked strength; this is not evidence that the served heads under-transmit.
Current source passes score-event conservation but is reconstruction-only. H-405 remains untestable on a historical point-in-time replay; no value credit.
Next: H-703 forecast owner: locate or capture pre-origin versions and first receipts for source games, adjudicate event residuals and 209 missing drive starts, then admit only… · Read finding and provenance →
STOP this preregistered version as underpowered before real labels; no real action-loss, D.ppd, G.margin, winner, causal policy or served-model gain is established. Preserve the H-451 shared-state mechanism work independently.
Next: H-451/INC-3932's existing construction owner can test a single legal inventory shared by action and clock, then grade its rules and real play/drive fidelity through E-01. A future… · Read finding and provenance →
Model disagreement with the open predicts the open-to-close move (CLV real, both halves, placebo null); margin ATS at the open is a lead (top-fifth 0.612, n=165) not an admitted gain; totals lose ATS.
Inconclusive by design: the scan survivor is real against its baseline and expected to be, and the existing served-head evidence puts the remaining off_pass EPA gap near a tenth of it with an interval spanning 0. The decisive test is the EPA-endpoint residual in would_see; it is not run here.
H-405 predictive claim untested. The exact current source prerequisite is a drive/event scoring view with historical source clocks and conservation, not another generic aggregate.
Next: H-703 is OPEN, owned by the loaded forecast lane. The lane picker selects it ahead of H-405 as a data_clock prerequisite. Once its source gate passes, forecast executes the… · Read finding and provenance →
MEASURED A CEILING, NOT THE FROZEN TEST: identity gain over position prior is 0.22-0.34% on first-new-team games, same as random game subsets and 3-4.5x below the 1% planning effect
DESCRIPTIVE REFERENCE, not a kernel test: the clock-only half-agnostic curve differs from real inventories by up to 0.56 timeouts (first half, last 3 min) and carries no spread; real pass rate falls 8-18 pts with own timeouts in the last 2 min. Confounded by clock inside window; no game effect measured.
H-648's premise as stated is not supported: the kernel issues giveaways in the same ORDER as as-of ball security (corr 0.173 [0.120, 0.228]; slope 0.060 [0.041, 0.080] giveaways per unit x), not backwards; the week-6 -0.32 was n=28 noise. Whether the slope is too SHALLOW is unresolved: real slope 0.150 [0.053, 0.253] (corr 0.085), ratio…
Next: (1) Level audit of kernel interceptions vs lost fumbles separately against real 2024-25 (kernel lost_fumbles node exists; add INT per dropback) to name which kind carries the… · Read finding and provenance →
def_pass (d_p) is the only channel named by the frozen rule: sim/close 0.71 pooled 2023-25 vs uniform kappa 0.89 (dev -0.18 [-0.25,-0.10]); below kappa in 2023 (-0.16), 2024 (-0.01, ~0), 2025 (-0.21); 2019-25 pooled 0.62 vs kappa 0.79. Run offence is the second candidate (0.65 pooled, dev -0.24 [-0.34,-0.08]; 0.50 over 2019-25) but 2024…