diff --git a/specs/STATE.md b/specs/STATE.md
index cde18fe..bdc6a73 100644
--- a/specs/STATE.md
+++ b/specs/STATE.md
@@ -1,6 +1,49 @@
 # VYNDR — STATE OF THE WORLD
 ### As of `cc8e478` (main, DEPLOYED + fingerprinted live), 2026-07-20. This file opens every future session. Update it when a train ships.
 
+## 🟢 RETENTION LIVE — model_snapshots capturing (2026-07-20, migration 025 applied)
+Phase 2 priority zero. **History now compounds from tonight.** Spec:
+`specs/snapshot-retention.md`.
+- **Why:** the ONLY model history was `ledger_entries` (640 rows / 6 game days)
+  and it stores NO model inputs. All ~30 other warehouse tables are empty. We
+  could score our own grades but could not replay a different model — the only
+  question a backtest answers.
+- **`model_snapshots`** (migration 025, APPLIED): append-only, ONE ROW PER GRADED
+  PROP PER SIDE PER CYCLE. Unique on `(snapshot_id, player_key, stat, line,
+  side)` so a retried cycle can't duplicate. RLS on, service-role writes only.
+- **VERIFIED CAPTURING (2026-07-20 03:02 UTC, 2 triggered cycles): 100 rows.**
+  MLB 50 (14 graded / **36 refused**), WNBA 50 (50 graded). `features`,
+  `grade_11`, `p_win`, `ev_pct` populated on 100% of graded rows; 1 model_version.
+- **Refusals are captured — the ledger drops them entirely.** First-ever refusal
+  breakdown: `juiced_no_edge` 18, `rare_event_over_below_line` 13,
+  `insufficient_data` 5. **The MLB gate refused 36 of 50 sides (72%)** — whether
+  that costs edge is now measurable for the first time.
+- Contract: retention is best-effort + injectable; `persist()` returns errors and
+  never throws. **A retention failure can never break a snapshot.**
+- 🔶 **EV LOOKS OVERCONFIDENT — do not surface it as truth yet.** First captured
+  EVs include +62.1%, +61%, +56.9% — implausible for real markets. Cause is the
+  quantile estimator clamping `p_win` at 0.95 (`PROB_CEIL`) off ~10 games.
+  **Hero v2 already ranks by highest `ev_pct`, so it will select the MOST
+  overconfident read.** Calibration (C-cal) must run on the harness before EV
+  drives anything user-facing.
+
+## 🔶 SETTLEMENT-CORRECTNESS AUDIT — two concrete findings (open, scheduled)
+Logged from the Phase 2 history report; do NOT fix inside a retention build.
+1. **ZERO pushes across 470 settled rows.** Implausible — hits/TB land exactly on
+   the number regularly. The exact-number push case is almost certainly
+   mis-scored (silently counted as hit or miss).
+2. **~28 props/day never settle** (Jul 17 MLB 86 graded/57 settled; Jul 18
+   103/75). Cause undiagnosed.
+Both are correctness bugs in the scoring that accuracy, ROI, CLV and every public
+record depend on.
+
+## 🔴 MODEL-VERSION CONTAMINATION in ledger_entries (permanent, mitigate by cutoff)
+`ledger_entries` mixes pre- and post-2026-07-19-fix grades with **no version
+marker**, and the eras cannot be separated retroactively. **Any backtest or
+accuracy claim computed off existing ledger history must treat the fix boundary
+as a hard cutoff.** `model_snapshots` stamps `model_version` + `code_sha` on every
+row so this can never recur.
+
 ## 🟢 BACKUP — DURABLE ON-BOX, VERIFIED BY READ-BACK (2026-07-20, deployed)
 The highest-severity open item is closed **on-box**. Off-box is a tracked follow-up.
 - **Runs as CODE, not a cron somebody must install** (`src/backupScheduler.js`,