project-decision-layer-audit — Findings

← All findingssource · memory/project_decision_layer_audit.md

11-agent read-only audit of the whole ecosystem, asking "how do we make decisions faster and flag bad plays". Conclusions, with the numbers that justify them.

Price is the only thing the ledger validates. Realized CLV<0 → n=603, 49.3% hit, −12.58% ROI, t=−3.44. CLV>0 → +0.47%, t=+0.26. Everything else is a price reader in costume: grade_v2's out-of-time Brier is 0.24581 vs 0.24590 for the raw market probability it consumes; its ROI is non-monotone and inverted at the top (A+ +2.15% / mean CLV −0.24¢ vs PASS +2.69% / +2.95¢). Across 30 model_matrix arms the best t-stat is +3.15 against a Bonferroni bar of 2.94 — nothing survives. Whole WNBA priced ledger +0.92% over 21,421; NBA −17.44% over 4,984.

Live bugs found (all still open unless noted):

is_live never set on player props in /odds/game — odds-service/app/db/repository.py:642 omits it, schemas.py:253 defaults False. /odds/props/bulk DOES carry it. Poisons data-hydrator's market_total → pace factor → every adjusted_projection.
Prod serves scout_ml as the WNBA list tier via an env override, contradicting nightly_picks.go:39-46 whose own comment calls it a measured loser (n=1,502 priced, 52.7%, −3.64%).
/odds/schedule 404s but clvcron/fanout.go:323 calls it → pre-tip fanout has never fired. That's why disciplined has 4 WNBA picks in its entire history.
[internal detail removed]
CLV is positive by construction: bet price is best-of-N (system_lineedges.go:334), close is Pinnacle-first (capture.go:347). Zero point is ~+1.0 to +1.3¢, not 0.
hit% and ROI printed over different denominators (system_lineedges 84.5% is over 20,413 decided, n_priced=331).
picks_manual.go:246 grades manual picks via EvalOneSidedOdds, whose 0.145 haircut exceeds the 0.0381 intercept → every manual pick gets StakeUnits=0 and value_flag can never be true. Fix before logging anything.
[internal detail removed]

DO NOT BUILD: another formula/matrix arm · grade_v2 as a selector or gate · a parlay optimizer (16 decided slips ever) · anything on /api/v1/bankroll (it assumes −110 on every bet, config.go:69) · a "fix disciplined so core fills" patch (Pinnacle prices 26% of the WNBA board — make formula membership a TAG on a price-ranked board instead) · chasing system_lineedges 84.5% (unbettable lines).

Do not read post-2026-07-07 improvement as a fix — unrelated formulas all improved together (slate effect); weekly pooled WNBA ROI swings ±16%.