docs: out-of-sample protocol, the two-year evidence and a ledger of adopted values

[yanivil]

Description

Implements #96. Documentation only.

• Wiki 03, new section "Out-of-sample check and the rule for changing a constant". States that every calibration to date was judged on the year the values were chosen on (the ablation ran on 63 sessions with 4 spec signals), then gives the two-year table from backtest run 34186512341 (--profile tuned --period 5y --days 500 --bars 500 --split 2025-09-09, 97 minutes): tuned +0.30 R on 167 signals in the unseen year against +0.28 on 159 in the tuning year; spec +0.53 against +0.24; legacy, from the matching local replay, +0.34 against +0.09. Per pattern: inverse H&S held, cups marginal, Wolfe negative out of sample. Four caveats: mostly bull years, today's constituents, time clustering, one exploratory pass over ten features.
• The rule. A candidate is chosen on one window and confirmed on another it never saw; it becomes the default only if the sign agrees in both windows and the pooled month-block interval excludes zero; slices under about 30 trades decide nothing; windows and run IDs are recorded.
• Ledger of adopted values with the runs that judged them, verified against the run logs: the tuned profile held; WW_TIME_SYM_TOL 0.45 failed its unseen window (Wolfe: negative out of sample after the +-45 % rhythm was chosen in-sample #102); the reward and patience thresholds were never tested alone; the cup trigger was judged in-sample only.
• Hypotheses under test (Breakout age at first report: rows first seen 2+ bars after the break ran +0.1 R against +0.45 for fresh ones #98 to Wolfe: negative out of sample after the +-45 % rhythm was chosen in-sample #102, deferred Per-signal probability and expected value: premature at about 120 resolved trades a year #97, enabling Point-in-time constituents for multi-year replays #95) and the note that the score gates but does not rank (Quality score does not rank outcomes #82), also added to the MIN_SCORE row.
• Wiki 03 --profile row now lists tuned; wiki Home mentions the split and intervals; README limitations state the out-of-sample status and the score's role; changelog.

The run matched the local replay of 2026-09-07 within a signal and a cent of R on every tuned and spec row; the legacy rows are labelled as local because the run replays spec, not legacy, alongside tuned.

Related Issues

Closes #96
Related: #91, #92, #93 (the tooling this page relies on), #95, #97, #98, #99, #100, #101, #102

Checklist

• Code follows existing project style and type annotations (no code changed)
• Threshold changes include a one-line "why" comment in scan.py (no thresholds changed)
• Offline unit tests pass (python -m pytest -q)
• Random-walk false-positive rate stays under 5% (detection logic untouched)
• Documentation updated in docs/wiki/
• CHANGELOG.md updated

🤖 Generated with Claude Code