Skip to content

v0.28.0 · Bounded report review

Choose a tag to compare

@hpoenixf hpoenixf released this 11 Nov 06:35
b34750d

Bounded report review

Review a saved report against its frozen source packet in a separate, persisted task. Findings retain evidence and uncertainty, the original report stays readable, and the CLI can inspect results from a fresh process. Review execution disables tools and does not automatically rewrite or republish reports.

This module builds on v0.26 Web evidence and v0.27 specialist profiles while retaining existing Runtime, Session, admission and budget owners. Dependency versions are unchanged.

Verification and limits

  • Fixed ten-task real-model cohort: 10/10 Runtime completions and readable Review records. Original report bytes remained unchanged; independent cold readback matched every result, and terminal replay made no new model calls.
  • The source packets are synthetic. This is Runtime qualification, not a financial-accuracy benchmark or a claim that every semantic verdict is correct. Verdict calibration, period inference and coverage limitations remain documented.
  • Local full suite: 1279 passed, 4 skipped, 0 failed. Historical unknown requests remain accounted for; they were not rewritten as successes.

中文

本版增加独立的报告复核模块:围绕已保存报告及冻结证据生成可追溯的复核记录,保留原报告,支持 CLI 冷启动读回。复核过程不调用工具,不自动改写或重新发布报告。

十个固定任务均完成真实模型 Runtime 验收,原报告与复核结果均可读取;材料为合成案例,不代表金融准确率测评。模型判断仍有校准与覆盖局限,不把回答完美作为 Runtime 发布条件。

Final package

Released from b34750d1515982dbaaaf4c1aa4da4c380173d753. Exact-main CI and PR CI passed. PR CI records 1279 passed, 4 skipped and 0 failed. The earlier CI failure and test-only clock/diagnostic repair remain documented in the qualification record.