DIME
Get the picksJoin

PROJECT · findings note

project_wnba_batch_misses_afternoon_games

The 21:00Z WNBA prop batch runs AFTER tip for ~21% of games (weekend afternoons). Verified 2026-08-06. Fix is a cron slot, but it's coupled to a risky CDK deploy.

← All findingssource · memory/project_wnba_batch_misses_afternoon_games.md

Verified independently 2026-08-06 against SDS /wnba/games/date/, comparing full tip timestamps (not naive time strings — 00:30/02:00 UTC are next-day evening games and are fine).

8 of the 39 remaining regular-season games (21%) tip at or before the 21:00Z prop batch:

  • 2026-08-08 — 2 games at 17:00Z, 19:30Z
  • 2026-08-09 — 3 games at 16:30Z, 19:00Z, 19:30Z
  • 2026-08-15 — 1 game at 17:00Z
  • 2026-08-16 — 2 games at 21:00Z (exactly at batch time)

Weekend afternoon games get their board generated after tip, or not at all. Independent corroboration from the news-latency agent over past slates: 181 of 2,683 props (~15/slate) affected, and 91.2% of those props already had a posted line by 14:00Z — so the data to produce them exists hours earlier. This is pure lost coverage, not a modelling issue.

Also a contamination risk, not just a coverage one: regenerating a slate after tip means the odds fetch can pick up in-play prices as if pregame — the exact poison documented in project_wnba_audit_2026_07 (bug 3).

BLOCKER — the fix is coupled to a deploy we do not want. The cron rules live in the same CDK stack as the batch DockerImageAsset, so infra/deploy-training.sh rebuilds and ships the image from current source. That would drag in the stranded semantics-v3 commits and fire the ratchet reset — see project_wnba_stranded_deploy. Three options:

  1. [internal detail removed]
  2. Deploy CDK from a tree pinned at the deployed commit aebe6a9 plus only the cron change.
  3. Wait for the offseason and do the full deploy properly.

Given the season ends 2026-08-19 (project_wnba_season_endgame), only options 1-2 recover any of the 8 games; option 3 concedes them.