Review of 2026-09-25T2224Z-h272-clv-ev-gate

Owner lane: Not recorded

What held

The direction and size of the vig-versus-CLV offset are replayed exactly: the breakeven mean CLV of about +0.386 pts (first grid k with pop EV > 0 at mean CLV 0.39), a mean EV of -4.95% at mean CLV -0.11, and the points gate's sampling half-width (about 1.96 x 1.85 / sqrt n) exceeding that breakeven below n of about 88 are all arithmetic that holds on the stored rows. The record froze the arms, reported all of them, disclosed the defective control and the overwritten first scaling run, and did not claim an edge.

What did not hold

(1) 'EXCLUDES ZERO BELOW under every conversion' rests on one wager (drop the -6.0 pt row and A1 spans zero) and on a fair-close assumption whose bias tolerance is about 0.5pp of cover probability; it is a margin of 0.49 percentage points on the upper bound. (2) 'CONFIRMED AND SHARPENED' with P4 and P6 missed and the operative consequence (scaling) exploratory. (3) The 0.23-0.84 figures are the best case over pop EV, not a rate at a stated EV. (4) The points-versus-EV contrast compares different nulls; the EV gate's <= 0.015 is its nominal size. (5) 'the process as run is below breakeven' is a statement about a 35-wager sample under an efficient-close assumption, not about the process; the CLV interval [-0.75,+0.52] straddles the breakeven +0.386, so the sample cannot say which side the process is on.

Finding and verdict

WEAKENED

Review scope

WEAKENED, not refuted. UPHELD: the zero of a points gate is not the zero of the bet: at the taken prices the breakeven mean CLV is about +0.39 pts under an efficient-close assumption, and a t95 lower bound on CLV points at n < ~88 can be cleared by a population that loses money; on the 35 stored rows every conversion has a negative point estimate. NOT UPHELD: the interval excluding zero below under every conversion (fails dropping the single largest |CLV| wager, and under a 0.5pp close bias), the sharpened scaling numbers as rates (maximum over a grid, post-freeze), 'CONFIRMED AND SHARPENED', and any reading that the process is below breakeven. The exact reproduction is refused; the replay was on a different snapshot.

Next action

(1) Bind the snapshot bytes (sha256) and the baseline release/model/entrypoint in the record before promotion; the original market-clv.duckdb is gone. (2) Test the efficient-close assumption on priced closes (odds_quote) across a season before the EV zero is used as a gate; state the tolerated close bias. (3) Any gate restatement should be reported at a stated pop EV (not the max over a grid) and with the wager-level unit replaced by game/week clusters. (4) A CHANGES-PROPOSED item for chain-report link 6 needs a separate PROMOTION record; this review is not one.