Cross-Ref Research

Use this skill for work under data/cross-ref/: adding or maintaining external evals, changing model mappings, regenerating published artifacts, auditing trust status, or updating the consolidated cross-eval report from generated facts.

Do not use this skill for ad hoc leaderboard commentary, fuzzy one-off model matching, or claims that will not be backed by source snapshots, mapping rows, generated artifacts, and runtime checks.

Start Here

1. Read data/cross-ref/README.md for the research workspace shape, artifact roles, and trust boundaries.
2. Read data/cross-ref/CONSOLIDATED_REPORT.md when the task touches published findings or unresolved mapping caveats.
3. Run run_cross_ref.py verify before trusting anything in data/cross-ref/results/: it checks whether the checked-in artifacts still correspond to current inputs, and exits non-zero when they do not.
4. Choose the narrow task path: output review, mapping row review, command execution, publication, or methodology change.
5. Work from the repository root. .venv is gitignored, so create it with uv sync if the checkout has none. Non-publish commands default to review outputs outside data/cross-ref/; use --publish only when intentionally updating checked-in generated artifacts.

Who Owns What

The script owns every derived number; the agent owns every judgment. See "Stage Ownership" in data/cross-ref/README.md for the stage-by-stage table.

• Agent-owned: mapping_status and llm_chess_player decisions plus their rationale/open_questions/evidence_refs, evals/*/SOURCE.md, CONSOLIDATED_REPORT.md, reviewing generated output, and the choice to --publish.
• Script-owned: normalization, score parsing, the mapping merge, status filtering, the join to elo_refined.csv, dedupe, all statistics, coverage and drop reasons, artifact writing, and provenance hashes.
• Everything you decide lands in a mapping CSV or a .md file. Nothing you decide is written into data/cross-ref/results/.

Procedure

1. Anchor the task in the smallest controlled surface: source snapshot, mapping CSV, adapter, generated result, report section, or failing command.
2. Preserve the workspace contract: source snapshots, mappings, mapping rationale, generated results, and code live in separate folders.
3. Treat model identity as research, in both directions. Map a model when the evidence is clear — a counterpart that exists in elo_refined.csv and fits the reasoning-effort rule belongs in the comparison. Hold a row as ambiguous, unmatched, or excluded when identity is genuinely uncertain; a mismatched or unclear reasoning effort is not grounds to hold, because effort resolves coverage-first by direction-aware nearest-tier substitution. The goal is the mapping that reflects the evidence: neither inventing matches nor withholding obvious ones.
4. Run the cheapest check that can disprove the current change before expanding scope. For mapping or adapter edits, that is usually a focused run_cross_ref.py command with explicit /tmp outputs or tests/test_cross_ref.py.
5. Publish generated artifacts only through data/cross-ref/run_cross_ref.py --publish. Do not hand-edit files under data/cross-ref/results/.
6. Update data/cross-ref/CONSOLIDATED_REPORT.md only from generated summaries and reports, then keep caveats explicit when unresolved rows constrain conclusions.
7. Before reporting done, run the task-relevant verification commands and the requested final checks. Use the QA handoff skill when the change is implementation-heavy or affects published artifacts.

Examples

Existing mapping correction:

1. Inspect data/cross-ref/mappings/{eval_id}.csv and mapping-research notes.
2. Update only rows with evidence; leave weak matches unresolved.
3. Run the eval in review mode or to explicit `/tmp` outputs, then tests/test_cross_ref.py.
4. If publishing, regenerate the eval artifacts, mapping_review.*, and cross-eval outputs as needed.

Existing output trust check:

1. Run audit in default review mode while investigating, and add `--publish` only when refreshing checked-in audit outputs.
2. Read audit status fields separately: reproducibility can pass while coverage remains review-needed.
3. Trace headline claims through *_summary.json before editing narrative reports.

Adding another eval:

1. Add evals/<eval-folder>/SOURCE.md with provenance, score meaning, columns, and caveats.
2. Add adapter and runner registration.
3. Add mappings/<eval_id>.csv with conservative statuses and rationale.
4. Add focused tests, generate /tmp artifacts, review coverage, then publish through the runner.

Refreshing an External Snapshot

Refreshing an existing eval's source snapshot with newer upstream data has its own steps beyond the examples above. Work through them in order.

1. Fetch upstream into a scratch file and keep the snapshot schema identical to the prior file. Record the canonical machine-readable URL in the eval's SOURCE.md so the next refresh starts from it (for ECI it is https://epoch.ai/data/eci_scores.csv, the published overall index; the leaderboard page renders that data dynamically and offers no direct download).
2. Re-key the mapping to the new rows. The mapping joins on (eval_row_id, eval_model_label), and eval_row_id is the row position assigned at normalize time, so a changed row set or order needs a fresh key. Index the existing mapping by eval_model_label, carry each retained model's reviewed decision onto its new position, and drop rows for models upstream no longer lists.
3. Reconcile the mapping with current data. Map new upstream models that have a clear LLM Chess counterpart, following the reasoning-effort rule in README.md: an exact effort match wins; unstated or unclear external effort assumes the highest (assume-highest); a stated effort absent from LLM Chess takes the nearest available tier in the same direction (nearest-tier), so high/xhigh/max go up, low/minimal go down, and medium breaks ties upward. Name the clause in reasoning_rule_applied. Revisit inherited mappings too, since a newly added LLM Chess model can be the better match — for example GPT-5.4 moves to gpt-5.4-high once that run exists. Keep a row unmatched when identity is genuinely uncertain or no counterpart exists, note why in open_questions, and raise true identity conflicts with the maintainer.
4. Update the snapshot name and its references. Snapshots are date-named, so rename to the new date and update the adapter SOURCE_PATH, SOURCE.md, the README.md artifact map, mapping-research/<eval>.md, and any filename assertions in tests/test_cross_ref.py.
5. Separate the data effect from any code change. Checked-in results/ baselines may predate the current code (signs: llm_chess_inputs.data_quality is null, or prediction.ols.in_sample is present). For a clean data-only diff, regenerate the baseline by running current code on the previous inputs — check the old snapshot and mapping out to /tmp — then diff the new run against that baseline.
6. Read the test results in context. tests/test_cross_ref.py pins dataset-derived counts and correlations, so a refresh will move several of them; refresh those expectations as part of the change and confirm the structural assertions still hold.

Mistake Metrics (wrong actions / wrong moves / mistakes)

Logs before 2025-03-16 underreported wrong actions and wrong moves, so the error/discipline metrics (wrong_actions_per_1000moves, wrong_moves_per_1000moves, mistakes_per_1000moves, and the player_wrong_* counts) are excluded from analysis by default. Do not quietly re-enable them.

• Each model row in elo_refined.csv carries min_game_date (earliest game start). A model is trustworthy for these metrics only when min_game_date >= 2025-03-16 — then every game is post-fix, so the published full-history value is already clean. The cutoff lives in one place: framework/data_quality.MISTAKE_STATS_TRUSTED_AFTER.
• To use these metrics in research, request clean mode: run_cross_ref.py <eval> --mistake-stats clean_only (or run_analysis(..., mistake_stats="clean_only")). It drops every model with min_game_date before the cutoff (or missing) and re-enables the repaired rate metrics for the remaining sample. It is research-only and refuses --publish; published artifacts always run with mistake_stats="excluded".
• Policy is drop, never recover: a model whose earliest game predates the cutoff is dropped whole, even if it also has later games. We do not recompute individual models from a post-cutoff subset.
• min_game_date is produced by data/get_refined_csv.py during the normal build; it is the only run-date provenance carried into the aggregate, so prefer it over model release dates for any "was this tested after X" question.

Optional Work Splitting

For large mechanical row review or command-output verification, an agent may ask another agent to inspect a bounded slice. Keep the request mechanical, provide exact files and acceptance criteria, and re-check the answer against the source artifacts yourself. These helper notes are not cross-ref artifacts and are not required workflow.

Edge Cases

• Python and pytest version strings are informational unless a command actually fails.
• The mapping is keyed by eval_row_id (the normalize-time row position) together with eval_model_label, so re-key it whenever a snapshot's rows change order or membership (see Refreshing an External Snapshot).
• A rerun-diff is a clean data comparison only when the baseline results/ were generated by the current code; regenerate the baseline first if they may be older.
• ARC COST (V3) remains unresolved in located official sources; keep cost interpretation conservative.
• Human baselines and benchmark-system rows stay visible in source and coverage outputs but are excluded from LLM Chess correlation samples.
• Release-controlled correlations are lower than raw Elo correlations; do not present raw correlations as model-capability proof without the timing caveat.
• If generated outputs differ after a rerun, inspect whether the difference comes from source, mapping, code, dependency behavior, or expected artifact metadata before publishing.
• The error/discipline metrics are excluded by default and only usable via --mistake-stats clean_only; see "Mistake Metrics" above. Never present them from a default run, and never publish a clean-only run.