Gate math = frozen-v1 bar. flagship comparison indicative (59-stock ML cache vs 503 broad); eqmom comparison clean (same universe/costs/grid).
BASE vs flagship — indicative, universe differs (overlap 4971d)
| leg | base | cand | | IS Sharpe | 1.02 | 0.89 | | IS maxDD | -30.8% | -54.7% | | OOS Sharpe | 1.95 | 1.74 | | OOS maxDD | -18.3% | -28.5% | IS p=0.465 -> NO; OOS Sharpe -> NO; OOS DD -> NO verdict: FAIL — target stands
BASE vs live-eqmom-stack — clean (overlap 5030d)
| leg | base | cand | | IS Sharpe | 0.86 | 0.88 | | IS maxDD | -48.0% | -54.7% | | OOS Sharpe | 1.80 | 1.74 | | OOS maxDD | -26.0% | -28.5% | IS p=0.301 -> NO; OOS Sharpe -> NO; OOS DD -> OK verdict: FAIL — target stands
Verdicts (2026-09-09)
Both live targets stand. Nothing in the 38-variant program beats them.
- Flagship wins soundly (IS 1.02 vs 0.89, DD -30.8% vs -54.7%; OOS ahead either way). The ML signal + 2.5xATR stops + 4xATR takes + profit lock + 15% caps + vol-target engine earns its keep — most visibly in drawdown control, which is where simple momentum bleeds. Caveats: 59-name curated ML cache vs 503 broad (indicative, not clean), and the cache's last 19 trading days were poor (full-own OOS 1.79 vs truncated 1.95) — worth watching, not a verdict. Survivorship fudge (1.5%/yr) applied per house methodology. Replication matches the P3 headline (0.93) to within annualization noise.
- Live-eqmom-stack edges BASE narrowly on identical ground (IS 0.86 vs 0.88 p=0.301; OOS 1.80 vs 1.74; DD better both windows). Resid+sector+ book+credit is worth small single-digit Sharpe bps and ~7pp of DD over raw — the opposite of expand2's no-book read, which shows the overlay carries the stack. Not significant; the stack keeps its job, BASE keeps pressuring it.
- Methodology repairs this round: eqmom baseline rebuilt on a data-driven calendar (188 market holidays were flat-zero padding diluting Sharpe 1.76->1.80); duel intersections verified day-for-day after two discrepancies (both resolved: holiday padding + flagship's bad tail).
Where a winner could still come from
- NM1 intermediate 12-7 vs THESE targets (only beaten BASE so far).
- Forward shadow (BASE live-decisions accruing; NM1 shadow candidate).
- A second ML-cache-style signal duel on shared names (needs PIT work).
- The live credit-gate fix (operator call) — changes live behavior, shadow first.