bench: commit curated-1 discrimination report pages (golden-curated-3)

[bborbe]

Summary

• Commit the four curated-1 discrimination report pages to bench/reports/, scored against the current committed golden (golden-curated-3, 158 entries)
• Models: opus (xhigh/full reference), deepseek-v4-flash-max, m3, m2.7 (all high/full) on the full 20-PR curated set
• Recall ladder: opus 0.826 (run 1) > deepseek 0.109 > m3 0.072 > m2.7 0.014 — a monotone model-strength ladder far beyond the measured noise floor (Opus 3σ ≈ 15.6% of mean)
• These pages are the reproducible artifact for the PR Review Bench discrimination result; they were previously claimed committed but none existed

Test plan

• Report pages render with correct Configuration blocks (model is the only differing field among the three high/full configs)
• Each page shows the full curated fixture scored (20/20 PRs, rows skipped 0)
• git status clean, no unrelated files

[ben-s-pull-request-reviewer]

Now I need to run the Step 4b-i jq filter manually. The changed files are all in bench/reports/ with .md extension. Let me evaluate the judgment-rule triggers against those paths.

Step 4b-i — Judgment-rule candidate computation:

The changed files from the diff stat:

• bench/reports/7b305997dc2b9b2f734f1499922286b7a8881ea5bd1967fad49275b83a8f1ba8.md
• bench/reports/d4558caff3dfefdff574d9a14345bb4007db9cb94a69c64e2e859c7d6e8b2dab.md
• bench/reports/d5a15f168212fe61bdf112b08c7a1787058cc5cd26a2a2d6c35a545465de79ca.md
• bench/reports/edb935210ee850e34cb521e7c50482923a65761bd4d66501c153a402f5f8f8eb.md

Checking each judgment rule's trigger globs against these paths:

• **/*.go → no Go files
• **/main.go → no main.go
• CHANGELOG.md → not CHANGELOG.md
• CLAUDE.md → not CLAUDE.md
• go.mod → not go.mod
• Makefile → not Makefile
• agents/**/*.md, commands/**/*.md, skills/**/*.md → paths are bench/reports/**/*.md, not agent/command/skill files

No judgment rules have triggers matching the changed files. Applicable set is empty.

Step 4c-sel CLASSIFY:

{"applicable": [], "skipped": {}}

Step 4d-sel ADJUDICATE: Since applicable set is empty, no rule-block reads needed.

Mechanical findings (pre-computed): 1 SHOULD finding from changelog/unreleased-entry-required.

Now checking for context-specific conventions — CHANGELOG.md is in the diff stat but the changed files are only under bench/reports/, so no context-specific doc needed.

Step 5 — Consolidated Report:

The mechanical funnel caught changelog/unreleased-entry-required as SHOULD because the repo has source changes (4 files added to bench/reports/) but no ## Unreleased bullet. This is a bench/fixture PR adding golden-curated-3 discrimination report pages — it is test-scoped infrastructure, not a source change that warrants a changelog entry. Classifying this as Nice to Have (Optional) with context.

---

{
"verdict": "approve",
"summary": "PR adds 4 benchmark report pages under bench/reports/ as golden-curated-3 fixture data. The single mechanical finding (missing changelog unreleased bullet) is a SHOULD that does not apply to bench/fixture-only PRs — no action needed.",
"comments": [
{
"file": "CHANGELOG.md",
"line": 1,
"severity": "nit",
"message": "changelog/unreleased-entry-required: repo has source changes (4 new files) but no '## Unreleased' bullet. For bench/fixture-only PRs this is a false positive — golden report pages are test infrastructure, not source changes requiring a changelog entry. *(rule: changelog/unreleased-entry-required)*"
}
],
"concerns_addressed": []
}