Verdict: Evidence-First Agent Harness for Reproducible Bug Fixes

Verdict is a system that demands concrete evidence before merging bug fixes. It treats bug investigation as an experimental process rather than a discussion. An LLM can quickly propose explanations for bugs, but actual reproduction is crucial. Verdict provides a structured approach to isolate the trigger, determine the failure rate, identify the control condition, specify repository ranges, and create regression tests to prevent recurrence.

The system consists of three agents that run sequentially: Hunter, Surgeon, and Insurance. Hunter searches for the trigger condition and records the failure rate under that condition. Surgeon localizes the suspect commit range or module boundary based on the trigger condition. Insurance converts the reproduction into a regression plan by creating a test case that reproduces the failure.

All runs are stored in a versioned ledger that includes command invocations, exit codes, stdout and stderr, timing, and snapshots. This keeps the evidence transparent and prevents manipulation. The system enforces strict boundaries, allowing only pre-approved commands, environment variables, file paths, and budget limits. It rejects any attempts to escalate privileges or perform unauthorized actions.

An agent cannot rewrite history or hide failed runs. Once a test case is created, it remains failing in the repository until a patch is developed. The maintainer reviews the evidence ledger before merging any test case. The agent never gains write access to the main branch. Observability features provide detailed logs of run ledgers, agent traces, budget consumption, and boundary violations.

Common failure modes include not reproducing the bug within the budget, which is recorded in the ledger for the maintainer to review. Verdict provides a rigorous, evidence-based method for handling bugs, ensuring that only reproducible fixes are merged into the codebase.