xingwudao/open-xquant/compare-experiments

Experiment Comparator

Use this skill when the user asks to compare two completed runs. The comparison is cross-experiment metadata, so do not write outputs inside either run directory. Resolve the comparison output root from .open-xquant/workspace.yaml when it exists: use paths.comparisons_dir for the comparison directory root and paths.comparison_registry for the summary registry. Fall back to comparisons/ and comparisons/comparisons.jsonl only when no workspace config value is present.

Do not use this skill in a version-governed workspace. Stop before reading run artifacts or writing comparison outputs and route the request to compare-strategy-versions, including when the user only says "compare two experiments". The version-aware comparator must validate each candidate against its claimed version manifest and backtest phase.

Classify the workspace exactly as the CLI does: it is version-governed if and only if workflow.layout == version_governed or the paths.versions_dir key is present. Otherwise it is legacy, even when .open-xquant/workspace.yaml exists. Use this legacy comparator only in the latter case.

Preconditions

Both runs must contain:

• strategy_spec.yaml
• metrics.json
• equity_curve.csv
• execution_assumptions.json
• research_bias_audit.json
• reproducibility_audit.json

If a run lacks metrics.json, say that the experiment has not completed a backtest and cannot be compared yet. If either run lacks audit or execution-assumption artifacts, stop and explain that audited comparison requires those artifacts before naming winners or writing spec-impact conclusions. If an audit contains fatal findings, do not present that run as comparable to an audited candidate without making the audit failure the primary result.

Workflow

1. Read both strategy_spec.yaml files and generate spec_diff.yaml.

• Compare fields recursively.
• Include path, run A value, run B value, and likely impact.

1. Read execution_assumptions.json and both audit artifacts for each run.

• Verify execution assumptions are comparable before interpreting metric

differences.

• Treat reproducibility or research-bias failures as blockers for

winner-style conclusions.

1. Read both metrics.json files and generate metrics_comparison.json.

• Include key metrics for each run.
• Include deltas and winners by return, Sharpe, and drawdown where present.

1. Generate figures under <comparisons_dir>/<comparison_id>/figures/.

• equity_overlay.png
• drawdown_overlay.png
• metrics_bar.png

Do not leave figures/ empty. If no figure will be generated, do not create the directory.

1. Write comparison_report.md.

• Explain which spec differences plausibly drove the metric differences.
• Do not claim causality when the evidence only supports association.

1. Append a summary row to <comparison_registry>.

Output Layout

With the default fallback paths:

comparisons/
├── comparisons.jsonl
└── <comparison_id>/
    ├── spec_diff.yaml
    ├── metrics_comparison.json
    ├── comparison_report.md
    └── figures/
        ├── equity_overlay.png
        ├── drawdown_overlay.png
        └── metrics_bar.png

Report Evidence Gate

When a compared input includes a current report package, validate its chart_build_result.json before using charts or report claims. Require the requested/applicable/generated/skipped set invariants, closed skip reason codes, the exact {path, sha256} manifest reference, and every generated asset hash. For each manifest asset, require a safe package-relative source.script, full lowercase source.script_sha256, and recompute the script SHA-256 from exact bytes. A script mutation blocks report-derived comparison evidence; a containing manifest hash or unchanged PNG does not excuse it.

Red Lines

• Do not modify either run's artifacts.
• Do not compare unaudited assumptions as if they were equivalent.
• Do not hide metric trade-offs by naming a single winner without context.