STATE: grading date binding fixed + blast radius quantified

2026-07-20 03:52:34 -04:00
parent 6415751f2e
commit b76e35f575
1 changed files with 60 additions and 0 deletions
@@ -133,6 +133,66 @@ exist locally; harmless, the data restores completely.

 ---

+## 🟢 GRADING DATE BINDING FIXED — wrong-opponent bug closed (2026-07-20 07:50 UTC)
+`computeFeatures.lookupTodayGame` → **`lookupGameOnDate`**: it now sends
+`?dates=YYYYMMDD` from the prop's BOUND game instead of taking ESPN's dateless
+"today". Grading, settlement, retention and the ledger finally reference the
+SAME game. **This was a MODEL-OUTPUT bug**: `opponentAbbr` feeds `opp_rank_stat`
+(a ±1.0 factor) and `isHome` feeds `home_away` (+0.5).
+
+**PROVEN against live ESPN, same instant:**
+```
+dateless "today"      CLE→PIT   NYY→LAD   LAD→NYY    (Jul 19 card)
+bound to 2026-07-20   CLE→MIN   NYY→PIT   LAD→PHI    (the real games)
+bound to 2026-07-19   CLE→PIT   NYY→LAD   LAD→NYY    (reproduces OLD exactly)
+```
+**Every opponent was wrong.** Measured at 07:5x UTC = **03:5x ET** — ESPN's
+dateless scoreboard was STILL serving the previous day's card, so **the affected
+window is wider than the 01:00/03:00 slots**.
+
+- **Honest degradation:** no bound date → NO fallback lookup. Records
+  `no_bound_game_date`, leaves `opponentAbbr`/`isHome`/`gameId` null, and engine1
+  omits those factors rather than scoring a wrong matchup. Tests lock both ways.
+- **Same class, also fixed:** Tank01 augmentation keyed its cache on TODAY's UTC
+  date; now uses the bound game date.
+- **Audited:** no other dateless/"today" lookups remain in the feature path
+  (weather is current-conditions by venue; park/pace are static).
+- Live cycles at 07:50/07:51 UTC wrote retention rows with `game_date 2026-07-20`
+  and `home_away` on 38/50 — games binding. MLB `opp_rank_stat` stays 0 for the
+  separate, known reason (ESPN's MLB team endpoint carries no defensive metric).
+
+### 🔴 BLAST RADIUS (report-only — gates the heal)
+**`ledger_entries` — grades by UTC hour (at-risk = late slots):**
+
+| UTC hour | grades | settled results | risk |
+|---|---|---|---|
+| 00 | 32 | 29 | 🔴 at risk |
+| 01 | 68 | 67 | 🔴 at risk |
+| 03 | 152 | 103 | 🔴 at risk |
+| 06 | 51 | 0 | 🔴 at risk (ESPN still on prior card at 02 ET) |
+| 13/14/19/22 | 390 | 379 | ✅ same-ET-day, correct |
+
+**At-risk: 303 grades (199 already settled with real results).** ~44 % of the
+public record was graded in the affected window.
+
+**`model_snapshots`:** 350 rows total — **50 carry `opp_rank_stat`** (the 03:02
+UTC WNBA cycle) and are the poisoned moat rows; 190 carry `home_away`. Rows
+captured 07:50+ are clean (post-fix).
+
+**Recoverable vs quarantine (PLAN ONLY, nothing written):**
+- **RECOVERABLE:** the correct opponent for a past date is still fetchable from
+  the date-pinned ESPN scoreboard, so a grade can be RE-DERIVED — but only where
+  the feature vector is reproducible. `model_snapshots` rows store the full
+  feature vector, so those are re-computable; **ledger rows do not** and would
+  need a full re-grade against historical inputs.
+- **QUARANTINE:** any at-risk ledger grade that cannot be re-derived should be
+  marked computed-against-wrong-opponent and EXCLUDED from harness/validation —
+  never silently trusted. It stays in the public record as a real settled bet
+  (the result happened) but must not train or validate anything.
+- **Scope note for the heal:** at-risk is defined by grading HOUR, not by proven
+  wrongness. Proving per-row wrongness requires re-deriving each row's opponent —
+  that IS the heal work.
+
 ## 🟢 game_date ROOT FIXED + 64 WRONG VOIDS REVERTED (2026-07-20 07:30 UTC)
 - **ROOT (not DNP, as first diagnosed):** PropLine emits **no `commence_time`**
  (grep-verified), so `ledgerService`'s `dateET(prop.game_time) ||