MLB Picks · Wednesday, August 19, 2026

# Correction — 2026-08-19 TDL picks

**Status**: DB corrected 2026-08-20. Original pre-game stamp preserved.

## What the stamp says

`v2/pipelines/mlb/tdl_picks/2026-08-19.json` (sha256
`de78e9d0ef95433920d0e9c8748b213c6b7bcd7b3ac2d8437d9250ca97359c86`,
stamped via OTS + Rekor at commit `8041f941` on 2026-08-19 13:45 UTC) recorded
2 picks:

- WSH F5 -0.5 (F5 spread, WSH @ TEX) — stake 45u @ 2.35
- SEA F5 -0.5 (F5 spread, SEA @ MIL) — stake 35u @ 2.35

Both showed model probability 0.5408, edge +11.53%, recorded 12:44-12:47 UTC.

## What actually happened

Both picks were **plateau artifacts** from an isotonic calibrator bug. On
2026-08-19, F5 training data was stale (game_targets.parquet had collapsed
to 2026-only rows) and the isotonic calibrator produced a monotone plateau
where multiple distinct raw probabilities all mapped to the same calibrated
value (0.5408). This produced identical "high-edge" tiles for both games on
the site's dashboard.

The operator clicked "took this pick" on SEA F5 -0.5 through the site UI
(persisted to DB as `tdl_picks` id 252 and `user_selections` id 267) before
realizing the edge was phantom. **The operator did not physically place either
bet at book.** WSH F5 -0.5 never made it to the DB — it appeared only in the
snapshot JSON due to a separate `_snapshot_tdl_picks` write racing with the
UI actions.

The only bet physically placed at book on 2026-08-19 was:

- **NYY ML @ BAL, decimal 1.86 (-116), $40 stake, FanDuel — WON 5-3**
  (`tdl_picks` id 253, recorded 19:44 UTC, `result: win`, `profit: $34.40`,
  CLV -2.22pp against 1.94 close)

## What was corrected

DB (2026-08-20 13:15 UTC via direct SQL):
- `DELETE FROM tdl_picks WHERE id=252` (SEA F5 -0.5 phantom)
- `DELETE FROM user_selections WHERE id=267` (matching tracker phantom)

Site performance page + tracker page (both DB-driven) now show the correct
1-0 day for 2026-08-19.

The **stamped JSON is intentionally unchanged**. Re-snapshotting the file
would break the pre-game cryptographic proof — the original OTS + Rekor
attestations remain valid for the original hash. This markdown is the
audit trail explaining the divergence between what was stamped and what
was placed.

## Root causes shipped

- `bbc0ef46` (2026-08-19): pybaseball_loader schema-shrinkage guard —
  prevents FG mlbapi fallback from corrupting the pitcher_stats cache
  (unrelated but same day's fire drill).
- `6e6e9bf7` (2026-08-19): `refresh_fg.py` + weekly reminder — durable
  Cloudflare bypass for FG data refreshes.
- `21c0df86` (2026-08-20): tracker.py pending count fix (push+PPD were
  miscounted as pending).

## Root causes still open

- Tracker `_pick_key` omits `side` — a tracked selection on one side of a
  market lights up "took this pick" on the opposite side after a refresh.
  Root cause of the "MIL F5 -0.5 phantom" the operator saw on 2026-08-19.
- F5 calibrator plateau — isotonic collapses raw probs to identical
  calibrated values on some days; sigmoid retrain attempted and reverted
  2026-08-19 after a validation harness regression. Deferred; documented as
  systemic in the 2-week audit (`plateau in `over_under` every single day`).