Write and audit the retrieval-workbench card

The final report explains one bounded experiment. It should make the arithmetic and judgment boundaries clear without turning fixture measurements into claims about users or future systems.

Goal

Write and audit project-root report.md as a concise retrieval-workbench card whose claims point to saved evidence and respect every interpretation limit.

Inputs

Use the manifest, term statistics, document weights and norms, query analyses and weights, score contributions, rankings, rank relevance traces, missing-relevant records, cutoff measures, paired differences, macro summaries, figures and plot data, read-back checks, and replay agreement or mismatch.

Deliverables

The card must explain in plain language:

• corpus, tokenizer, query, and index identity and the exact searchable-term rule;
• the smoothed natural-log idf, sparse document weights, norms, query weights, and zero-term or zero-norm behavior;
• the integer term-count baseline and tf–idf cosine contribution formulas;
• candidate selection, raw versus twelve-place ranking scores, and all ranking tie stages;
• complete judgment coverage, rank-level cumulative hits, and missing relevant documents;
• fixed-cutoff prefix sizes, precision, recall, F1, and every denominator;
• candidate-minus-baseline differences, unweighted macro means, artifacts, plot scales, read-back checks, and replay evidence; and
• what the fixture cannot establish about relevance, users, significance, fairness, causality, representation, or future performance.

Include one small worked path from a query through contributions, a ranked document, a judgment, and a cutoff record. Distinguish observations from interpretation and name any first mismatch honestly.

Checks

Audit every statement against a saved record or explicit supplied rule. Reject wording that claims semantic understanding, causal effect, statistical significance, fairness, representative judgments, or future performance. Check that no score is called a probability, that complete fixture coverage is not presented as ordinary search coverage, that zero denominators remain explicit, and that macro means are not pooled measures.

Check the card names both systems, exact rounding and ties, positive-ranking behavior, missing relevant documents, all cutoff rules, paired differences, shared plot scales, artifact names, read-back, and replay outcome. A reader should be able to locate the evidence behind every numerical claim.

Workspace

Keep the card at the project root as report.md. Do not replace the report with a leaderboard, trophy, gauge, or decorative results page. Preserve all machine-readable artifacts and figures beside it.

Hints

Review

Give the card to someone who has not opened the source code. Can they reproduce one score, explain a tie, find a missing relevant document, recompute a cutoff measure, and tell whether replay agreed? Mark every sentence that outruns the artifacts.

How to check your work

Checks compare report.md with the evidence-limited reference card after your own audit. The supplied fixture explains formulas, judgments, denominators, and limits; it does not hide uncertainty behind a leaderboard.