H51 verdicts: both fades KILLED at taker-floor bars; 51b maker-proxy (Sharpe 0.90, corr 0.01) noted as possible H52

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: martex-dev

PROJECT_MEMORY.md

@@ -3,7 +3,7 @@
 The knowledge file: ledger, results, meta-findings, lessons, open
 assumptions. PROJECT_STATE.md = what runs now; this = why and what we know.
 
-## Trial ledger: 112 run + 2 registered (H51) = 114. Every new spec raises
+## Trial ledger: 114 pre-registered trials, all run. Every new spec raises
 ## the DSR bar for everything. Do not test without a numbered doc FIRST.
 
 ## Hypothesis ledger (docs/hypotheses/, docs/research/)
@@ -48,7 +48,7 @@ assumptions. PROJECT_STATE.md = what runs now; this = why and what we know.
 | 42b | Rotation + chandelier stop | **CANDIDATE — beats champion on all metrics** | Sharpe 1.47 vs 1.10, MDD -29% vs -58%, prop 73.0%>62.8% @0.5x, **DSR 0.992 > 0.95 bar** (104 trials); paper account #4 since 2026-07-12 |
 | 43 | Combo batch on rot-stop base | screen: only bounce admits; **43a KILLED (eval bars)** | rot-stop x V1 corr 0.521, x rotation 0.821 (blends dead); rot-stop+bounce Sharpe 1.55, +79%/yr, DSR 1.000 but prop 47.5%<73%, MDD worse -> replaces H41 as own-capital archive |
 | 44-50 | Retail intraday batch (maker regime, 15m Bybit) | ALL 7 claims KILLED; **H44 ORB + H45 first-hour INVERTED (significant)** | fade earns 0.16-0.20%/event = 4-5x maker toll; sessions/funding/levels/bursts/VWAP noise -> H51 |
-| 51 | Intraday fade strategies (51a fade-ORB, 51b fade-1st-hour) | REGISTERED, not yet run | bars: net Sharpe>0.7 + corr<0.30 vs rot-stop (diversifier), taker-floor costs |
+| 51 | Intraday fade strategies (51a fade-ORB, 51b fade-1st-hour) | BOTH KILLED (taker floor) | 51a Sharpe 0.14; 51b 0.44 < 0.7 bar; maker-PROXY 51b 0.90 corr +0.01 (most independent stream ever measured) — true maker-fill model = possible H52, queued behind sprint |
 
 ## Prop-firm simulation results (real CFD rules, 20k paths, EOD approx)
 

PROJECT_STATE.md

@@ -109,15 +109,15 @@ config/secrets/hyro.json (gitignored).
 4. Paper shakedown continues untouched (4 accounts nightly); Jul 25
    gate now only governs the PATIENT fallback if all 3 sprint attempts
    bust.
-5. TRACK 2 COMPLETE at info level (2026-07-12): H44-50 all retail
-   claims killed; **H44 ORB and H45 first-hour momentum significantly
-   INVERTED** — intraday crypto REVERTS (meta-finding 7). Fade effects
-   clear the maker toll 4-5x. **H51 pre-registered** (51a fade-ORB,
-   51b fade-first-hour, strategy grade, taker-floor costs, diversifier
-   bars vs rotation-stop): NEXT session's build+run. Survivors join
-   the book only via runbook (5th paper account / funded stage),
-   never mid-attempt. Data: data/intraday/ (12 Bybit perps, 15m,
-   2021+, ~2M bars).
+5. TRACK 2 COMPLETE (2026-07-12): H44-50 all retail claims killed;
+   ORB + first-hour momentum significantly INVERTED (intraday crypto
+   REVERTS — meta-finding). H51 fade strategies BOTH KILLED at the
+   registered taker-floor bars (51a Sharpe 0.14, 51b 0.44). Loose end:
+   51b maker-PROXY Sharpe 0.90 at corr +0.01 to rotation-stop — a true
+   maker-fill engine model (limit orders, missable fills) is the one
+   remaining intraday question (possible H52), queued BEHIND the
+   sprint build. Data: data/intraday/ (12 Bybit perps, 15m, 2021+).
+   Ledger: 114.
 3. Backlog (docs/research/backlog.md): options/Deribit VRP data project,
    correlation-spike de-risking, carry infra (post-eval, own-capital);
    own-capital book = rotation-stop + crash-bounce overlay (43a:

docs/hypotheses/51-intraday-fade.md

@@ -1,7 +1,8 @@
 # Hypothesis 51 — Intraday Fade Strategies (from H44/H45 inversions)
 
-Status: **PRE-REGISTERED (2026-07-12)** — no test has run yet.
-Trial ledger: +2 -> 114.
+Status: **BOTH KILLED at the registered taker-floor bars (2026-07-12).**
+Trial ledger: +2 -> 114. Verdicts at the bottom; one honest loose end
+(true maker-fill model) noted for a possible H52.
 
 ## Stated reason (per process)
 
@@ -52,3 +53,26 @@ inversion joins H16's archive (real info, unusable as strategy).
 - Effect discovered and strategy tested on the same history (family-
   selection caveat, same as H42); a paper record guards deployment.
 - Bybit perp data starts 2021 — no 2017-2020 regimes in this panel.
+
+## Verdicts (2026-07-12, scripts/h51_fade_study.py, 1,989 days)
+
+- **51a fade-ORB: KILLED.** Taker floor Sharpe 0.14 (CAGR -0.8%).
+  The 0.16%/event info edge dies against the ~0.13% real round trip.
+  Maker-proxy sensitivity: Sharpe 0.59 — still under the bar.
+- **51b fade-first-hour: KILLED as registered.** Taker floor Sharpe
+  0.44 < 0.7 (bar 1 explicitly named the taker-floor case). The
+  maker-proxy run printed Sharpe 0.90, corr +0.01 and the script's
+  generic bar labeled it CANDIDATE — that label is NOT a verdict: the
+  proxy discounts fees but still assumes every entry fills. The doc's
+  true maker variant (limit fills that can be MISSED — and fades miss
+  their best entries) was not actually modeled; the truth lies between
+  0.44 and 0.90.
+
+## Loose end (possible H52, NOT yet registered)
+
+corr +0.01 to rotation-stop makes 51b the most independent return
+stream ever measured here — IF the edge survives honest maker-fill
+modeling. Graduating requires: engine limit-order support with
+crossed-range fill logic, then a new pre-registered trial. Queued
+BEHIND the sprint build; register only with the engine extension in
+hand. Everything else in the intraday family is closed.