Review of 2026-09-23T0325Z-h239-preflip-join-census

Owner lane: Not recorded

Reproduction

REFUSED -- source record binds no protocol, runner, input or result hashes (experiment_state EXECUTION, protocol_sha256/runner_sha256/result_sha256 None, input_artifacts []). Not retrofitted. The source record's own script and output bytes were, however, checked: h239_join_coverage.py and .out.json hash to the values in its provenance block (832d865687..., a1444e2387...).

Finding and verdict

WEAKENED

Review scope

The enumeration (which tables, which writers, coverage, exposure fractions) survives an attempt to refute it at the current tree. Two secondary claims do not: the served-0-is-out-of-support characterisation (0.0 is an existing ~1% sentinel) and the ~09-29 timing (the selector can flip on the first played week-3 row). Source record remains UNADMITTED (no evidence contract); this is a REVIEW record, not a PROMOTION and not an outcome-blindness attestation.

Next action

(1) If the price of the flip matters, it is due now, not 09-29: any played week-3 row in player_week flips latest_rows. (2) A fitted-arm question, not answered here: does a tree split on coach_pos_* treat 0.0 as the sentinel bucket (178 rows) coherently, and what does serving 100% of rows there cost -- forecast lane, walk-forward, with the coach pair held at its 2025 value as the free baseline. (3) The writer gap remains a build item (regime.build/build_kalman, a rel_coach writer, drives.build_volume on the reingest path).