diff --git a/specs/STATE.md b/specs/STATE.md
index 87c3681..2501268 100644
--- a/specs/STATE.md
+++ b/specs/STATE.md
@@ -1,5 +1,5 @@
 # VYNDR — STATE OF THE WORLD
-### As of `9f60ceb` (main, DEPLOYED), 2026-07-21. This file opens every future session. **Start with the CURRENT STATUS + OPEN ITEMS block below.**
+### As of `6dd6f59` (main, DEPLOYED), 2026-07-21. This file opens every future session. **Start with the CURRENT STATUS + OPEN ITEMS block below.**
 
 ---
 
@@ -131,6 +131,73 @@ exist locally; harmless, the data restores completely.
 `specs/snapshot-retention.md` (Phase 2 report + schema) ·
 `specs/audit-data/grade-collapse.md` · `specs/audit-data/gate-simulation.md`.
 
+---
+
+# 🟡 SESSION 76 — PLATOON SPLITS (regressed) (2026-07-21)
+**Built + tested + deployed. Not reaching live props** (same plumbing gap).
+
+## Gates — both clear; one was ALREADY closed
+1. **Splits** = statsapi pull, 1 call/hitter:
+   `people/{id}/stats?stats=statSplits&sitCodes=vl,vr&group=hitting`.
+2. **🔴 CORRECTION — batter handedness is NOT pending.** S69's Tier-A join did
+   it: `statcast_aggregates.bats` = **604/604** (210 L · 327 R · 67 S). The S69
+   STATE note said pending; the data says done.
+3. **Point-in-time holds** if fetched before first pitch (a split queried this
+   afternoon can't contain tonight). A HISTORICAL backtest would use
+   season-final and leak → **clean measurement is forward-accruing.**
+
+## 🔑 THE SPINE — regression IS the feature
+`regressed = (PA·observed + K·prior) / (PA + K)`, **K = 600 PA**, prior = the
+hitter's **OWN** blended rate (not the league's). The question is *"is he
+DIFFERENT vs this hand than he normally is"* → his own line is the null, so a
+hitter with no evidence gets **nothing**.
+Weights: `30 PA → 4.8% · 130 → 17% · 400 → 40% · 1000 → 63%`.
+
+### The make-or-break test (both halves, tested)
+| Same .310 vs LHP | weight | multiplier | δ |
+|---|---|---|---|
+| **30 PA** | 4.8% | 1.011 | **0.003** |
+| **400 PA** | 40% | 1.096 | **0.023** |
+**8× apart.** A test asserts the ratio stays **> 5** — if the regression ever
+breaks, the suite fails instead of projections quietly drifting onto noise.
+
+## Real data — most hitters get NOTHING, and that's correct
+| Player | vs L | vs R | vs LHP mult |
+|---|---|---|---|
+| Josh Bell (S) | .259 / 126 PA | .248 / 265 PA | **1.005** (17% wt) |
+| Aaron Judge (R) | .246 / 76 PA | .248 / 185 PA | **0.999** (11% wt) |
+Bell *looks* like a platoon splitter until the sample speaks. Neither is
+material. **Expect most hitters to get ~nothing — honest output, not failure.**
+
+## Honest-absent — 5 distinct routes, all exactly 1.0
+`batter_hand_absent` · `pitcher_hand_absent` · `splits_absent` ·
+`not_applicable` (stat) · missing SIDE → falls back to prior (weight 0), never 0.
+
+## INDEPENDENT of the environment — deliberate
+Park+weather compose into ONE coefficient (both describe the stadium). Platoon
+rides its **own** `matchup` slot with its own label — entangling them would make
+both **harder to attribute** when the instrument scores them. Mirrored on under,
+capped ±15%, **inverted for strikeouts** (higher K rate → higher K prop).
+
+## 🔴 THE PLUMBING GAP — now FOUR orders old
+`snapshotService` still doesn't attach **venue + first-pitch + opposing-pitcher
+hand** per prop and pass `environment` / `matchup` into `attachChallenger`.
+**Park, weather AND platoon all adjust NOTHING live.** Everything else is ready.
+
+## 🔴 And still: no settled row carries `p_win`
+All seven adjusters remain unmeasurable.
+```sql
+select count(p_win) champion, count(p_win_challenger) challenger
+from ledger_entries where user_id is null;
+```
+
+## Recommendation (repeated, now urgent)
+**Stop adding adjusters.** The next order should be ONLY: (a) attach venue +
+first-pitch + opposing SP hand to props, (b) pass environment/matchup through,
+(c) confirm `p_win` lands on a settled row. The measurement stack is now far
+deeper than the plumbing feeding it.
+
+
 ---
 
 # 🟡 SESSION 75 — WEATHER MODULATION (2026-07-21)