STATE: Session 71 — challenger deployed; verification gap documented honestly

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCNgGSt5qvcLxaeQqa7Zpj

specs/STATE.md

@@ -1,5 +1,5 @@
 # VYNDR — STATE OF THE WORLD
-### As of `474ebc5` (main, DEPLOYED + fingerprinted live), 2026-07-21. This file opens every future session. **Start with the CURRENT STATUS + OPEN ITEMS block below.**
+### As of `f2da9dd` (main, DEPLOYED), 2026-07-21. This file opens every future session. **Start with the CURRENT STATUS + OPEN ITEMS block below.**
 
 ---
 
@@ -131,6 +131,72 @@ 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 71 — LAYER 3 STEP 2: CHALLENGER PROJECTION (2026-07-21)
+**Code-complete + deployed + unit-proven. NOT yet observed writing a live row —
+see the verification gap below.** Champion serves users UNCHANGED.
+
+## What it is
+`src/services/challengerProjection.js` — the champion's `p_win` **adjusted** by
+the Layer-2 axis vector in **log-odds space** (a nudge can't cross 0/1 and means
+the same at p=.5 and p=.9). Total capped at **0.45 log-odds** — a lean, not a
+re-forecast. Every deviation is attributable: `challenger_adjustments` stores the
+named axis + signed nudge. Only mechanically obvious stat↔axis pairs are mapped.
+
+**Identical to champion by construction** where there's no signal — unremarkable
+player, thin sample, unmapped stat, or no classification all return the champion
+value byte-for-byte with a stated `reason`. The A/B therefore differs ONLY where
+archetype-awareness could help or hurt.
+
+**Isolation:** `adjust()` is pure; the champion field is read, never written; the
+served snapshot payload is still `enriched` (untouched); a challenger throw is
+caught so it can never break the pipeline it's measured inside. Statcast loads
+ONCE per snapshot run — grade-time I/O stays at zero.
+
+## Induced on real players (local, real data)
+| Prop | Champion → Challenger | Via |
+|---|---|---|
+| Judge HR **over** | 0.42 → **0.447** | BOMBER +.22, WHIFF RISK −.11 |
+| Judge HR **under** | 0.58 → **0.553** | exact mirror |
+| Judge strikeouts | 0.55 → **0.55 (δ=0)** | WHIFF RISK +.11 / GRINDER −.11 **cancel** |
+| Skubal Ks over | 0.60 → **0.702** | WHIFF, TRAPDOOR, CANNON aligned |
+| Skubal hits-allowed | 0.50 → **0.392** | a K arm allows fewer hits |
+| Bell / 12-PA sample | **unchanged** | `no_distinctive_axis` / `thin` |
+
+## 🔴 VERIFICATION GAP — read before assuming this is recording
+**Zero live rows carry `p_win` or `p_win_challenger`** (`latest_lock` still
+`2026-07-21 02:58:45`). Three compounding reasons:
+1. **Immutability refuses to backfill.** Every prop on the current slate was
+   locked BEFORE these columns existed; `ignoreDuplicates` correctly skips them.
+   **First live rows arrive on the next NEW lock** (new game date / new line).
+2. **`ledgerWritten` OVER-REPORTS.** It counts `chunk.length`, not actual
+   inserts — 14 induced snapshots all reported "ledger 25" while inserting
+   **zero**. This misled my own verification; worth a one-line fix.
+3. **`statcast_aggregates` is MLB-ONLY (1,354 rows; 0 wnba/nba).** I induced on
+   WNBA, where the challenger correctly no-ops for lack of mechanism data.
+   **The challenger currently applies to MLB only.**
+
+**Verify with:**
+```sql
+select count(p_win) champion, count(p_win_challenger) challenger,
+       count(*) filter (where challenger_delta <> 0) moved
+from ledger_entries where user_id is null;
+```
+Expect first non-zero after the next MLB slate locks.
+
+## HELD
+**Promotion** (challenger → serving) and **any superiority claim** stay held
+until the ledger proves it on segmented volume. Baseline volume: blended n≈500
+≈ 5 days; per-segment n≈200 ≈ 2–4 weeks.
+
+## Next
+1. Confirm first live champion+challenger rows after the next MLB lock.
+2. Fix `ledgerWritten` to report real inserts.
+3. Extend mechanism data beyond MLB (or accept MLB-only challenger).
+4. Let volume accrue; then compare Brier/calibration segmented.
+
+
 ---
 
 # 🟢 SESSION 70 — LAYER 3 STEP 0+1: THE MEASUREMENT INSTRUMENT (2026-07-21)