vyndr/specs/window-bug-class-audit.md

The window-bug class, hunted — three more paths, and the forward re-audit rule

PHASE 0 — the audit

The defect class: a fixed short window used AS the season/base rate. getStatRows is the single path feeding meta.gameLogs, which is where estimateProbability derives its base rate, so every branch of it is a base-rate path. The feature builders are the second surface, because l20_avg is the season reference projectionFor reads.

path	window	classification
getStatRows MLB → estimator base rate	fullLog	CORRECT (fixed 929fd81)
mlbGameLogFeatures → l5/l10/l20_avg, l10_stddev	last10 = 10	DEFECTIVE
espnStatsAdapter.parseGameLog → NBA/WNBA logs	rows.slice(0, 20) = 20	DEFECTIVE
getStatRows NBA/WNBA ESPN branch	inherits the 20-cap	DEFECTIVE (via source)
getStatRows NBA/WNBA Python branch	getGameLogs(..., 20)	DEFECTIVE-but-dormant (service offline in prod)
pitcher engine (pitcherEngine, skillProjection)	reads statcast profiles, no game log	N/A
pitcher props (strikeouts) via getStatRows MLB	fullLog	CORRECT — fixed by the same change
settleSource	already reads the full log (S64)	CORRECT
playerIntelService, streaksService	display/streak surfaces, not forecasts	N/A

The pitcher answer matters and is good news: pitcher props run through the same getStatRows MLB branch, so 929fd81 repaired them too — there is no separate defective pitcher base-rate path.

The one that was hiding in plain sight

mlbGameLogFeatures carries this comment:

l20 = all available (the season per-game reference projectionFor needs)

Built from last10, l20_avg was a ten-game average wearing a season label — and it feeds both the consistency (cv) pull inside the estimator and projectionFor, which decides refusals. Same class as the base-rate bug, same file, and it survived the previous repair because that fix touched only getStatRows.

---

PHASE 1 — fixes

path	fix	API cost
mlbGameLogFeatures	read fullLog, fall back to last10	ZERO — same response
espnStatsAdapter.parseGameLog	drop the slice(0, 20) cap	ZERO — same payload, already parsed
NBA/WNBA Python branch	left as-is	service offline in prod; fixing it would be speculative

No new API calls anywhere. Both fixes widen data that was already fetched and then discarded — the same shape as the original repair.

Measurement status, stated honestly

These are serving changes for the MLB feature path and the NBA/WNBA log path. Their before/after resolution is NOT measured here, and deliberately: the only way to measure it today would be to reconstruct the repaired forecast over old rows, which is the reconstruction-vs-served trap this order explicitly refuses. They ship as code fixes with the measurement deferred to accrual, which is the honest sequencing.

---

PHASE 2 — the forward re-audit rule, in code

MODEL_VERSION is bumped to engine1@2026-08-07-fullwindow, so every snapshot from this commit forward is self-identifying. retentionService already stamps it onto model_snapshots, so no new plumbing was needed.

model/reAuditEligibility.js encodes the rule:

• isEligible(row) — true only for rows carrying the repaired marker.
• assess(rows) — counts eligible DATES, not rows, because dates have been the binding scarcity in every interval this session.
• ACCRUAL (frozen) — pre-stated minimum dates per measurement:

measurement	minimum eligible dates
calibration re-fit	10
hits factor lift	10
prior verdict re-audit	14
rbi lineup-slot gate	14

A test locks the case that would otherwise be invisible: a MIXED table of 330 rows where only 30 carry the new marker returns eligible_dates: 3, not 330 rows' worth of false confidence. Once both generations sit in the same table, a naive count would happily fit a map on a blend of two different forecasters.

---

PHASE 3 — the honest board

What happened: the champion computed its season rate over ten games. Found by resolution decomposition, not by a test failing. Fixed in two lines. It no longer loses to a frequency table — it beats it CI-confirmed only on total_bases, ties on rbi and runs, and leads on the hits point estimate.

Consequences, each labelled:

• CALIBRATION — WITHDRAWN. CALIBRATION_DEPLOYED is empty. Maps were fitted on the retired forecast. Re-fits on repaired-champion settled rows. Waiting on accrual: 10 dates. Not to be refit on reconstructions.
• FACTOR VERDICTS — SUSPECT. Every prior null and every THEATER was measured against a champion worse than a frequency table; signal added to noise reads as noise. Re-audit on accrued rows. Direction UNKNOWN — some may pass, some may still fail. Not pre-priced. Waiting: 14 dates.
• HITS FACTOR LIFT (1.39%) — UN-REMEASURABLE. Needs rows produced by the repaired champion. Waiting: 10 dates. The factors remain wired and transmitting (43f65d3); only the lift number is unquantified.
• RBI LINEUP-SLOT — RE-QUEUED. Lands after the champion is sound and rows accrue. Waiting: 14 dates.

Pre-registered re-audit order (each runs only when its bar is met):

1. Re-fit calibration on repaired-champion rows (10 dates)
2. Re-measure hits factor lift (10 dates)
3. Re-audit prior factor verdicts (14 dates)
4. Run rbi lineup-slot through the two-part gate (14 dates)

Then STOP and accrue. Nothing further can be honestly measured until the board fills with rows the repaired champion produced.

---

Invariants

No measurement on reconstructions — hard refusal, and it is why Phase 1 ships code without numbers. Serving-path changes by design for the MLB feature path and NBA/WNBA logs; frozen model modules verified unchanged. p_win never mutated. No Bonferroni slot — base-rate repair and a bug hunt, not causal factors.