Review of 2026-09-23T1438Z-h200-involved-week-board

Owner lane: Not recorded

Reproduction

REFUSED. Source record input_artifacts is [] and protocol_sha256 is None; the as-run snapshot (h200, 1676 MB) is gone and the fresh one is 2081 MB; the unpinned runner now reads season 2026 (450 rows) instead of 2025 (633) because 2026 has reached MIN_LEAGUE_WEEKS=3. Bound by hash and unchanged: runner sha 45faa8a4..., result out.json sha a3cf102d..., source record sha 0639cbe7.... playerweek/playermodel.py (eeeb6a3c...) last changed e0a0441c, 2026-09-22, before the run; git diff HEAD is clean for playermodel.py and share.py.

Finding and verdict

UPHELD

Review scope

Narrow claim only: at HEAD f6511555 latest_rows(INVOLVED_WEEK=True) selects a week with a player_pbp_l4 row whenever the player has one at or before the selected week (0 stranded on the season-2025 board, 633 rows, and on today's season-2026 board, 450 rows); INVOLVED_WEEK=False strands 115 (2025) / 33 (2026). NOT upheld as stated in the record's note and NOT evidence that the served board no longer carries hard-zero pbp features: 90 of 633 rows (64 with no pbp row + 26 with a pbp row whose l4 columns are all NULL) still reach the model as exact zeros; the record measures key presence, not value.

Next action

If the served hard-zero question is revisited: measure the 90 exact-zero rows at value level (the 26 NULL-valued pbp rows and the 8 cap-excluded players) against their realized involvement, since the record's key-presence invariant cannot see them.