AUDIT-REPLAY-20260910 — same-condition replay before and after the execution-engine fix
- Run date: 2026-09-10
- Execution model:
claude-opus-5 / high - Audit document: external audit
STOCKBACK_AUDIT_2026-09-10.md(base commit1f5a37f) - Purpose: check whether the engine fix changes earlier verdicts. Not a search for returns.
| generated_at | 2026-09-10T00:00:00+09:00 |
Replay contract
With the same data, signals, costs, capital and parameters, only the execution model was switched between v1 (existing) and v2 (audit fix) and run side by side. No cost normalization or parameter changes were mixed in. The basket is the preregistered 30 symbols of EXP-011, and the 1st cell is locked_oos × normalized_actual_proxy × research_unhalted.
Result 1 — Turtle squeeze (ATR 1.5)
| Contract | v1 CAGR | v2 CAGR | v1 MDD | v2 MDD | v1 trades | v2 trades |
|---|---|---|---|---|---|---|
turtle_only | -1.89% | -1.89% | -32.58% | -32.58% | 299 | 299 |
turtle_squeeze_60 | -1.37% | -1.37% | -15.18% | -15.18% | 89 | 89 |
turtle_squeeze_125 | -1.22% | -1.15% | -12.10% | -12.10% | 78 | 78 |
turtle_squeeze_250 | -0.85% | -0.66% | -12.30% | -12.30% | 78 | 78 |
The grounds for the EXP-011 verdict did not change. The key figures, 9.74% for the 7-symbol control vs -1.70% for the 23 new symbols, are identical on both engines. The rejected verdict stands.
Result 2 — Connors RSI(2) daily
| Variant | v1 CAGR | v2 CAGR | Difference | v1 trades | v2 trades |
|---|---|---|---|---|---|
baseline | 3.39% | 3.39% | +0.00% | 331 | 331 |
hard_stop_7pct | 0.17% | 0.17% | +0.00% | 361 | 365 |
time_exit_5_bars | 1.18% | 1.18% | +0.00% | 357 | 357 |
hard_stop_7pct_and_time_exit_5_bars | 0.07% | -0.24% | -0.31% | 373 | 378 |
As the audit predicted, the base form without stops is completely identical. A-01 triggers only when a scheduled open exit competes with a same-bar stop, so it does not touch the hard_stop_pct=None path. Only the variant using both a stop and a time exit worsened by -0.31%, a result of A-02 (restoring entry-bar stops) lowering returns.
Interpretation
Items whose performance improved after the fix are not interpreted as new alpha. For turtle_squeeze_250, the +0.19%p is an accounting correction from A-07's handling of take-profits on opening gaps, not a strategy improvement. Likewise, the -0.31%p deterioration is not hidden.
There are three reasons the impact is small.
- Current candidates have low turnover (1~10 OOS trades per symbol), so the bars where the bugs would trigger are rare.
- The basket is the 30 most liquid symbols in the train period, so 0-volume sessions (A-03) are almost absent.
- The main contracts do not use take-profit targets (
target_r_multiple=None), so A-07's main path is inactive.
A small impact does not mean the bugs were trivial. For higher-turnover contracts, small caps and contracts using take-profit targets, the impact could be much larger. In particular, 5-minute contracts have hundreds of times higher turnover, so the impact of A-04, A-05 and A-06 cannot be estimated from this replay.
Not run
- The 5-minute (EXP-006/009) replay was not run. Those contracts are already
rejected, and measuring the impact of the A-04~A-06 fixes requires a separate run. - The full-universe sweep (1,042 symbols) was not rerun on v2. The existing sweep results are
v1outputs and are marked as such in their reports.
Source and migration evidence
Source: reports/experiments/AUDIT-REPLAY-20260910.md
Original SHA-256: 502f2225f432fe89bf0a011156dd5267f10f7248ff13f0705f544fe8866a043e
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.