E-14. Harness 1st run — all 4,752 combinations (2026-07-14)
- Run: run_id=260714_full (+smoke1 100), BTC 1h train 21,128 bars → val comparison. 0 failures. Cumulative ledger of 4,852 attempts (DSR denominator). Pass line preregistered under D-017.
- Result: Stage 0 (trades ≥30) 4,279 → Stage 1 (DSR>0.5) 0 → Stage 2 0. 0 survivors.
- Distribution: eligible annual Sharpe median -1.08 (costs eat almost everything), best +1.73 vs null E[max] 3.65 — even 1st place is half the noise ceiling. The failure was not close; it was overwhelming.
- Block patterns: long-only alone averaged -0.42 (beta to BTC's upward drift in the train period), trend-type entries (donchian/squeeze) were less bad than mean-reversion types, and the top 10 were all long — not an edge but the shadow of market beta [inference]. short/both averaged -1.3~-1.4.
- Why E[max] 3.65 > the calibration prediction of 2.2: the real-ledger Sharpe dispersion (annual sd≈0.96) exceeds pure estimation noise (≈0.64) — genuine heterogeneity between combinations (different cost structures) mixes into the variance. This is a known property of the Bailey prescription; the gate behaves more conservatively than registered [inference].
- Verdict: on single-symbol BTC 1h, the search space of block grammar v1 has no edge exceeding costs. The 6 manual classes + 4,752 automatic combinations reach the same conclusion — mutually confirmed.
Source and migration evidence
Source: backtest/docs/experiment-log.md#E-14
Original SHA-256: 069c4aaaf53a20930ca8d94ba6365b4ce2b6bc560171d662a4c179ea55a6857c
Raw data and account information are not part of the public package. The original results do not replace statistical or scientific revalidation. The English text is a translation; numbers, tables and charts are carried over unchanged.