Instruction file imported from chickenopoulos/thequantgpt-wrapper (.cursor/rules/45-selection-log.mdc). Copyright stays with the author.

Selection log

• After any backtest, PSA, or tqg_alpha_mine.py run, artifacts/selection.json must exist (the harness writes it).
• Quote IS Sharpe, OOS Sharpe, N, k, DSR together. Never IS Sharpe alone. For formulaic mining also quote IS/OOS Rank IC; OOS Rank IC is n/a until --reveal-oos.
• Quote Lo SE / 95% CI and MinBTL when sharpe_diagnostics is present. Quote PBO only from artifacts/pbo.json / selection; if the PSA return matrix is missing, say PBO is not computable.
• Changing the economic rule → new run_id and --related-to the prior run (family N carries forward).
• Picking a PSA “best” cell as the new canonical spec → N includes the grid. Representative/stable cell is the robustness output, not a silent re-baseline.
• Combining signals or averaging winners → increment k; list strategy_spec.signals (or ensemble_weights / ranking.sub_scores).
• Inline candidate loops must call record_trials(n, kind="inline_grid"). Each newly scored alpha formula is one trial.
• report.md is rules + metrics. Economic narrative is labeled interpretation, not evidence.
• Do not present IS − DSR or IS − null as a live Sharpe forecast. DSR and the null measure manufactured IS Sharpe.
• Do not fail packaging because DSR is low. Fail only if the selection log cannot be written.