audit/README.md

Reproducibility Audit

This directory documents the data provenance for every figure, table, and quantitative claim in the paper. It is intended as a reviewer-facing reproducibility manifest: any number, panel, or claim in the paper should be traceable to a specific backtest run, CSV, or source document.

Audit anchor

• Commit: a4bf9e6 (fix(figure): use 6M case-study runs for US sector matrix)
• Audit date: 2026-04-28

The audit was performed after fixing a previously-discovered run-ID mismatch in the U.S. sector matrix figure (the script wired the figure to GPT-5.4 3M ablation runs but the paper text described it as a 6M case study). All other figures, tables, and prose claims were re-verified against actual run metrics during the same audit.

Files in this directory

File	Scope
README.md	This overview
run_inventory.md	All backtest runs referenced anywhere in the paper, indexed by experiment family
figures.md	Per-figure provenance (which script, which runs, which data file)
tables.md	Per-table provenance plus verified cell values
claims.md	Key quantitative claims in the prose with data sources

How to verify a number in the paper

1. Locate the figure/table/claim in question.
2. Look it up in figures.md, tables.md, or claims.md.
3. The audit doc names the source (run ID, CSV path, or source doc).
4. The actual run artifacts live in the parent repository under reports/backtest/<run_id>/ (metrics.json, trades.csv, equity_curve.csv).

Metric scope (v1)

Each runs/**/metrics.json ships the metric set actually used in the paper plus the supporting fidelity diagnostics:

• Performance: total_return, annualized_return, max_drawdown, max_drawdown_duration, volatility, sharpe_ratio, sortino_ratio
• Benchmark anchor: benchmark_total_return, benchmark_annualized_return, benchmark_source
• Fidelity: total_trades, trading_days, avg_position_days, avg_cash_ratio, avg_gross_exposure, win_rate
• Profile-specific: e.g. value_filter_pass_rate, value_consistency_score for Fundamental Value

Benchmark-relative fields (alpha, beta, tracking_error, information_ratio, excess_return, calmar_ratio, up_capture, down_capture) are intentionally not shipped in v1. They were computed per run during development but require a harmonized benchmark-return series and a risk-free rate that are out of scope for the matched-window protocol used here, and the implementation gave different values across the U.S. and CN paths during validation. A v1.1 release will reintroduce them after cross-substrate harmonization; the v1 paper does not depend on these fields.

The same applies to derived/all_metrics.csv and derived/gpt54_robustness/gpt54_us_3m_summary.csv: the v1 release ships only the columns above. Bundle maintainers can rerun tools/strip_benchmark_relative_metrics.py to enforce this scope on a freshly built bundle.

Data layout (parent repository)

<repo_root>/
├── latex/                          # Paper source (this is the Overleaf project)
│   ├── sec/                        # Main-text sections
│   ├── appendix/                   # Appendices
│   ├── figures/                    # Compiled figure PDFs/PNGs
│   ├── scripts/                    # Figure/table generation scripts
│   ├── data/artifact_ingest/          # Optional static input artifacts for local rebuilds
│   └── audit/                      # ← This directory
└── reports/
    ├── backtest/<run_id>/          # Per-run artifacts (metrics.json, trades.csv, ...)
    └── multi_personality/<bundle>/ # Cross-mandate comparison bundles