[tzhouam]

JiuwenSwarm’s knowledge recognition and file coverage do not measure PR-review accuracy. Add a Chinese comparison of original author documents and the merged PR #290 knowledge base, plus a reproducible isolated GLM‑5.3 A/B evaluation.

The study freezes source/document hashes and actual PR merge bases, runs 12 PRs × two contexts × three repeats through subscribed Zcode with 13 workers, and uses separate Codex source auditing and anonymous scoring. Closed read-only tools enforce source boundaries and the cumulative two-page/6,000-character document budget. Complete native traces and a 3,612-file hash index remain outside Git; the checked-in JSON is about 70 KB.

Results: 72 terminal trials, 55 valid reviews, 12 successful independent scoring calls. On six comparable primary PRs, conditional precision is 69.12% for author docs versus 76.76% for the knowledge base. Primary completion falls from 21/24 to 15/24; paired native time averages 426.33 versus 416.77 seconds. Rate limits and protocol failures prevent a causal acceleration or overall-benefit claim. Recall is limited to two primary PRs with independently confirmed issues; four older forks are exploratory.

Validation: 3,033 local tests passed, 17 skipped; 135 evaluation fixtures, knowledge/wiki validation, documentation links/citations/SPEC, CLI/doctor and installed-wheel smoke checks. Independent code/report audits passed. Code CI passed before report publication; final delivery CI is checked on this PR.

Report: eval/knowledge-depth/jiuwenswarm-original-docs-comparison-cn-20261003.md.
Reproduction: eval/jiuwenswarm_ab_README.md.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.