diagnosing-matching-eval

name: diagnosing-matching-eval description: Use when investigating why the Index Network opportunity evaluator scores matches wrong, or when asked to explain, diagnose, or report on matching-eval results — turning a matching-eval run report (packages/protocol/eval/matching/runs/*.json) into a root-cause and fix report.

Diagnosing Matching-Eval Failures

Overview

The matching eval harness scores the opportunity evaluator (invokeEntityBundle) against a golden corpus. A run report (--report → eval/matching/runs/<ts>.json) records, per case per run, each candidate's actual score, role, and the evaluator's own verbatim reasoning. This skill turns that artifact into a report that explains why the evaluator fails and what to change.

Core principle: The reasoning string is the evidence. Every claim about why a case fails cites the evaluator's own words; every proposed fix cites a verified file:line. No guessing, no memory.

Get a run report first

• From packages/protocol: bun run eval:matching -- --report (add --runs 7 to cut noise). Writes eval/matching/runs/<timestamp>.json (gitignored). Analyze the newest one.
• Do NOT analyze baselines/matching.baseline.json — reasoning is stripped from it. It has scores only. You need a --report artifact for the "why".

Procedure

1. Read the reasoning, not just the scores. For each failing or flaky case, read every run's candidates[].reasoning alongside the assertion detail lines. The reasoning is the model explaining itself — that is your evidence.

2. Classify the cause before prescribing a fix. For each failure, decide which it is and say so:

   • Evaluator weakness — the reasoning is wrong, OR correct reasoning produces a wrong score/role/match.
   • Corpus / harness miscalibration — the expectation is unsatisfiable as written. The runner surfaces any candidate scoring ≥ minScore (matching.runner.ts:26, currently 30) and the scorer counts a surfaced candidate as matched (matched = score > 0, matching.scorer.ts:30). So if a case sets match:false but a scoreBand whose max is ≥ minScore, the band and the match expectation cannot both pass — that is a corpus bug; fix the case, not the prompt.
   • Do not default to "fix the prompt." Name the classification with evidence.

3. Separate signal from noise. Three runs is a small sample. An all-fail (0/N) or all-pass (N/N) case is conclusive as-is. But a mixed result is flaky — before drawing any conclusion from it, re-run just that rule to see where it settles: bun run eval:matching -- --rule <rule> --runs 7 --report. Flaky (mixed passes across runs) is not a regression (the harness's own baseline diff). Never call sample-to-sample variance a regression.

4. Locate the responsible clause precisely — and in the RIGHT prompt. The evaluator has two system prompts in src/opportunity/opportunity.evaluator.ts:

   • systemPrompt (~line 21) — strict, single-pair, ≥70 threshold.
   • entityBundleSystemPrompt (~line 77) — permissive, batch, ≥30 threshold.

   The eval calls invokeEntityBundle, which uses entityBundleSystemPrompt only. A fix applied to systemPrompt changes nothing in the eval. grep the rule, read it, cite the exact file:line, and confirm it lives in entityBundleSystemPrompt. Check whether the same rule is duplicated elsewhere. Attribute behavior to the right component: surfacing threshold = runner minScore; matched = scorer (score > 0); scoring/role rules = the prompt.

5. Write the report, then verify every quoted clause and line number by opening the file — never paraphrase a prompt rule from memory.

Report format

• Header: model, runs/case, aggregate pass-rate, path to the analyzed artifact.
• Per failing rule: what the case expects → what happened each run (quote the candidate's reasoning) → cause classification (weakness vs miscalibration, with evidence) → fix with a verified file:line in the correct prompt.
• Order fixes by impact. Mark each fix's confidence, and whether a higher---runs re-run confirmed the failure is systematic rather than flaky.

Common mistakes

Red flags — stop

• About to recommend a prompt edit without having opened the prompt file.
• You cited a line range, not a line — or didn't confirm which of the two prompts it's in.
• You called something a regression off a single run report.
• A "why it fails" claim has no quote from the candidate's reasoning.