Bugs Are Innocent Until Reproduced: Building Verdict, an Evidence-First Agent Harness

说明与用途

Verdict 将 GitHub 问题转化为可复现的调查，记录每一次运行并生成可独立验证的证据。

问题

Flaky bug reports often end as "cannot reproduce" or patches that cannot be proven fixed, leading to uncertainty and wasted effort.

结果

Verdict produced a reproducible record for TrueForge issue #417, demonstrating 10/10 failures under a stalled endpoint and 0/10 under a responsive endpoint, with a verifiable hash and a draft PR awaiting maintainer approval.

工作流

• Create a GitHub issue for the bug
• Configure Verdict with approved commands, knobs, and budgets
• Hunter searches the condition matrix and records all runs
• Surgeon narrows the reproduced condition to the smallest suspect range
• Insurance converts the reproduction into a regression plan and draft PR
• Maintainer reviews the evidence and approves publication
• Verifier re-runs the evidence to confirm reproducibility

证据

Verdict turns a GitHub issue into a bounded investigation. It runs an approved command repeatedly under approved conditions, keeps every observation and refuses to claim a reproduction unless the evidence crosses a deterministic threshold.

来源

The verifier returns: { "verdict": "REPRODUCED", "stalledRuns": 10, "responsiveControls": 10, ... }

来源