zenith-lowtf — Can we beat V5? (3-lever improvement study)

15m engine. Levers: correlation gate · exit asymmetry · range-fade entry. Pre-registered, phased IS→matrix→OOS. Conservative costs. OOS run once at frozen winner.

Out-of-sample — winner vs V5 (the decisive test)

Config	Trades	Win%	PF	MAR	Return	MaxDD	t
V5 baseline	315	61.27%	1.825	3.103	60.36%	-8.07%	3.79
Winner	300	61.0%	1.85	2.958	58.01%	-8.17%	3.8

Winner config: {'corr_max_per_group': 1}

Phase 1 — per-lever contribution (IS, BTC solo)

Lever	Trades	Win%	PF	MAR	Return	MaxDD	t
baseline (V5)	86	55.81%	1.647	0.618	11.59%	-4.18%	1.68
Correlation gate	86	55.81%	1.647	0.618	11.59%	-4.18%	1.68
Exit asymmetry	112	32.14%	1.64	0.547	7.47%	-3.09%	1.45
Range-fade entry	129	56.59%	1.427	0.325	9.86%	-6.8%	1.42

Survivors carried to matrix: ['A_corr', 'B_exit']

Multiple-testing discipline (Deflated Sharpe)

Total distinct configs evaluated: 10 · Expected best-of-N t under the null: 2.146 · Winner OOS t: 3.8 · Survives haircut: True

Verdict gate (pre-committed): VERDICT: NO significant improvement | OOS MAR 2.958 vs 3.103 (+10%? False) | PF 1.85 vs 1.825 (True) | t 3.8>=2 (True) | trials=10

Integrity — All levers default-OFF; baseline reproduces V5 bit-exact (86 tr, PF 1.647 on 15m BTC IS). Range-fade fake-profit bug (counter-trend stop using trend htfLine) caught and fixed before the run. Look-ahead discipline preserved (fades reuse shifted pivot columns). OOS touched once at the frozen winner; never used to select or tweak. Every config logged for the trial count above.