E-17. [gpt 5.6 sol] Maker-cost counterfactual audit of existing strategies (2026-07-15)
- Scope: all of RSI 81 + vol breakout 36 + cascade 12 + 5m shock 16, harness 4,752 full proxy screen, then 301 exact replays. A total of 446 were replayed under four cost scenarios.
- Classification: already positive as taker 209 / newly recovered with maker entry 36 / maker on both sides only 19 / negative even with no costs 86 / not recovered 96.
- val comparison of the 36 newly recovered: 5 exploratory survivors. val maker Sharpe 0.62~1.38, PF 1.20~1.53, 52~79 trades. 4 were also positive on val taker with Sharpe 0.74~1.13, but their train taker was below 0.5, so they were weak candidates under the existing gate.
- Frequency: 52~79 trades over 293 val days = 0.18~0.27 per day. Short of the several-per-day target.
- Limits: an upper bound on costs that keeps every existing next-open fill. It does not reflect unfilled limit orders, queue position or adverse selection, so it is not grounds for reviving an official strategy.
- Verdict (D-021): the value of building a real limit/OCO engine is confirmed. The 5 survivors are carried forward only as a reference group.
Source and migration evidence
Source: backtest/docs/experiment-log.md#E-17
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.