jsboige · Sep 12, 2026
diff --git a/...I.Notebooks/QuantConnect/ML-Training-Pipeline/docs/L5_vol_targeted_composite.md b/...I.Notebooks/QuantConnect/ML-Training-Pipeline/docs/L5_vol_targeted_composite.md
@@ -35,7 +35,12 @@ allocation + execution gate) **does not beat S4 v2 alone**.
 | **L5 full** (trend + VT + gate) | 0.613 | **-0.236** |
 | no-VT (trend + gate, no vol-targeting) | 0.589 | -0.260 |
 
-Per-seed (L5 full vs S4 v2): 0 -> -0.338 | 1 -> -0.267 | 7 -> +0.040 | 42 -> -0.380.
+| Seed | Δ Sharpe |
+|------|----------|
+| 0 | -0.338 |
+| 1 | -0.267 |
+| 7 | +0.040 |
+| 42 | -0.380 |
 
 ## Key Findings
 

diff --git a/MyIA.AI.Notebooks/QuantConnect/ML-Training-Pipeline/docs/M11c_DM_SIGNIFICANCE.md b/MyIA.AI.Notebooks/QuantConnect/ML-Training-Pipeline/docs/M11c_DM_SIGNIFICANCE.md
@@ -71,7 +71,7 @@ $\Pr(\hat S_a^* - \hat S_b^* - (\hat S_a - \hat S_b) \geq \hat S_a - \hat S_b)$
 | XRP-USD h=20 kelly_har_mu60 | 0.0731 | 0.0700 | concordant |
 | DOT-USD h=20 kelly_har_mu60 | 0.0849 | 0.0970 | concordant |
 
-All 5 |Δp| ≤ 0.02. The LW2008 closed-form and the stationary bootstrap converge
+All 5 \|Δp\| ≤ 0.02. The LW2008 closed-form and the stationary bootstrap converge
 to essentially the same p-value, so the LW2008 result is **NOT** an artefact of
 the asymptotic-normal assumption — it reflects genuine uncertainty given T and
 the cross-correlation between active and passive returns.

diff --git a/MyIA.AI.Notebooks/QuantConnect/projects/FuturesTrend/README.md b/MyIA.AI.Notebooks/QuantConnect/projects/FuturesTrend/README.md
@@ -214,27 +214,34 @@ quand la cible est proche de la position actuelle.
 ### Statut courant (c.1111, lane `myia-po-2027:CoursIA-2`)
 
 - Le code **compile statiquement** (`ast.parse` PASS, 7 fonctions / 1 classe / 465
+
   lignes, EOL LF, 0 secret literal).
 - **Aucun backtest exécuté** : le verdict SOTA est `RECOVERABLE-MACHINE` (credentials
+
   QC absents sur po-2027 — vérifié firsthand `env | grep -iE "QC_|QUANTCONNECT"` =
   0 hit). La jambe QC Cloud (compile/backtests) sera déléguée à une lane CoursIA-2
   équipée, sur cette branche, **sans transmission de secret** (Tell secrets-hygiene
   règle 1 : jamais de clair sur dashboard/PR/commit, `os.getenv("KEY","<literal>")`
   interdit).
 - **Verdict futur** : `BEATS` / `NO BEATS` / `INCONCLUSIVE` selon Sharpe/CAGR/MaxDD/
+
   PSR/exposition/coûts/ordres sur fenêtre >= 2016-2026, **sans présumer** du
   Sharpe 0,944 vs 0,749 rapporté par l'article #15989 sur 2020-2023 (fenêtre
   favorable non-représentative).
 - **REPAIR c.1107** : deux défauts détectés par le préflight adjoint po-2025
+
   (`msg-20260911T040615-4c08xy`) avant lancement des runs QC Cloud — (a) carry
   proxy identique à EWMAC(8,32), (b) FDM clip à l'inverse de la docstring.
 - **c.1063 increment** : scalaire per-horizon `sqrt(slow/32)`, warmup 2x cohérent,
+
   history bulk 1 call. Détails dans le commit `e41a4eb8d632`.
 - **REPAIR-3 c.1109** : FDM requalifié en breadth multiplier (Tell c.1069 strict
+
   honnêteté référentielle) + retarget delta direct (pas d'aller-retour fabriqué).
   Préflight adjoint po-2025 `msg-20260911T043805-i7tl0g` ; jambe QC po-2026
   suspendue jusqu'au nouveau head exact.
 - **REPAIR-5 c.1111** : docstring `_breadth_multiplier` sign-invariant magnitude
+
   concentration (Tell c.1069 strict honnêteté référentielle — l'`abs()` efface les
   signes, `[10,10] == [10,-10]`) + `config.json` Carver FDM stale → breadth
   multiplier sign-invariant. Préflight adjoint po-2025
@@ -243,11 +250,13 @@ quand la cible est proche de la position actuelle.
   jusqu'à cette dissipation par le worker (REPAIR-6 c.1111 = cohérence prose
   README ↔ source).
 - **REPAIR-6 c.1111 (worker)** : cohérence README ↔ source REPAIR-5 — section
+
   « Note Tell c.1069 strict — FDM requalifié » étendue avec le défaut sign-invariant
   (la prose c.1109 disait « bonus quand le book est unidirectionnel », faux) +
   statut courant étendu avec REPAIR-5. Amend borné scope unique
   `FuturesTrend/README.md`. Push `--force-with-lease=refs/heads/feature/15549-carver13-futures:5c316080f10ffcf0dd011a4da1a7ce77dad20277`.
 - **REPAIR-7 c.1113 (worker)** : correction sémantique « effective breadth /
+
   inverse concentration » (Tell c.1069 strict honnêteté référentielle, par
   préflight adjoint po-2025 habilité n°3 urne `delivered` Tell c.15069 strict,
   `msg-20260911T063424-qnc0q9`) — la formule `sum(|f|)/sqrt(sum(f^2))` mesure

diff --git a/docs/qc/qc-comparative-backtests.md b/docs/qc/qc-comparative-backtests.md
@@ -369,6 +369,8 @@ Standardized backtest results from QC Cloud via MCP qc-mcp-lite. Period: 2018-01
 | Framework_Composite_FamaFrenchAllWeather | 28882145 | 2018-2025 | 0.338 | 6.578 | 13.1 | 22.9 | `e9ac7c66` | COMP framework FamaFrench+AllWeather (FF20/AW80; FamaFrenchAlpha risk-adj momentum top-2 of VLUE/MTUM/SIZE/QUAL/USMV quarterly NO SMA200 + AllWeatherAlpha static SPY/IEF/GLD/XLP monthly, MultiStrategyPCM 20/80 monthly, NullRisk + ImmediateExecution, IBKR margin). Sharpe 0.338 = PERIOD-OVERFIT COLLAPSE: headline 0.588 (2010-2026 sweep) + OOS 0.684/PSR 87.5% (bt `b08c8956`, 835 dates) do NOT survive alignment — 0.684→0.338 (-51%), PSR 87.5→22.9% = small-OOS-sample artifact not robust edge; catalog 2015-2025 (bt `70415edc`) 0.472. The 80% AllWeather sleeve drags composite BELOW standalone FamaFrench 0.445 (#72) on aligned window BUT buys tight MaxDD 13.1% (among tighter backbones). Corrects the "FamaFrenchAllWeather 0.684 leader" refs in findings #33/#34 (that 0.684 was OOS-only). Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
 | Framework_Composite_EMATrend | 28911253 | 2018-2025 | 0.611 | 16.670 | 27.9 | 19.783 | `3095a263d5bd30df181ec002c0a52b72` | COMP framework EMA-Cross+TrendStocks (EMA70/Trend30 sweep WINNER; EMACrossAlpha 20/50 EMA on 5 Mag7 AAPL/MSFT/GOOGL/AMZN/NVDA + TrendStocksAlpha double-confirmation Price>SMA200+EMA20>EMA50 on 15 mega-caps incl the 5 Mag7, MultiStrategyPCM weekly, NullRisk + ImmediateExecution, IBKR margin). Sharpe 0.611 = SURVIVES alignment with a mild drop (catalog 0.741→0.611, -18%), NOT a period-overfit collapse (contrast #90 FamaFrenchAllWeather 0.588→0.338). HIGHEST-SHARPE COMP backbone on the aligned window (edges c2 0.574 #34, FamaFrenchAllWeather 0.338 #90, c1 0.258 #33); CAGR 16.67% = highest no-ML backbone. MAG7 SURVIVORSHIP CAVEAT: EMA sleeve 100% Mag7 → Sharpe partly a Mag7-decade artifact (loses pre-2018 ramp + absorbs 2022 drawdown on alignment); highest Sharpe ≠ most robust constitution — c2 0.574 (factor-diversified across 25 stocks) is the more defensible COMP leader constitution-wise, EMATrend is the higher-Sharpe Mag7-concentrated one. G.1: repo main.py had drifted to EMA40/Trend60 "starting point"; converged to the 70/30 deployed winner (matches QC Cloud 28911253 + catalog lines 36/119) for an apples-to-apples aligned run. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
 
+| Project | QC ID | Period | Sharpe | CAGR% | MaxDD% | PSR% | Backtest ID | Notes |
+|---------|-------|--------|--------|-------|--------|------|-------------|-------|
 | RiskParity (inverse-vol aligned) | 33286158 | 2018-2024 | 0.361 | 8.304 | 18.4 | 11.3 | `0eaf6f5b` | RISK true inverse-volatility risk-parity (#46 catalog; rank-based RP v4 docstring "BEST": SMA200 trend filter + 12m-momentum rank × close/std60 inv-vol weighting, BND safe-haven <2 trending, monthly; SPY/EFA/GLD/DBC/TLT+BND, IBKR margin). Aligned 0.361 = -30% vs student 2015-2025 0.514 (proj 31872286 *non-aligned) = regime-4 fee-friction (bonds+commodities NOT fee-homogeneous, AllWeather -30% pattern). Catalog 0.40→0.361. Best risk-parity-family member (cf Cloud-RiskParity-Composite 0.027 #23 mis-named equal-weight). PSR 11.3% non-significant. Promoted Tier 4→2. totalOrders=0 = wrapper artifact (CAGR 8.3% ⇒ real trades) |
 | DualMomentum (Antonacci aligned) | 33286487 | 2018-2024 | 0.350 | 8.413 | 14.9 | 9.6 | `b1cb13c9` | IND Antonacci canonical dual momentum (#48 catalog; absolute filter price>SMA200 AND 6m>0 + relative 12m ranking all-candidates momentum-tilt, BND safe-haven <2 trending, monthly; SPY/EFA/EEM/TLT/GLD/DBC+BND, IBKR margin). Aligned 0.350 = catalog 0.35 FLAT (no fee-collapse); tighter MaxDD 14.9% than sibling RiskParity 18.4% (dual-filter + BND refuge). PSR 9.6% non-significant — student 0.493/PSR 54.9 (proj 31798582, 2yr) was a small-sample artifact (finding #35 FamaFrenchAllWeather pattern). Promoted Tier 4→2. totalOrders=0 = wrapper artifact (CAGR 8.4% ⇒ real trades) |