verification-before-completion

To stop agents from claiming a bug is fixed without testing, enforce evidence-based verification protocols that require running specific validation commands and reporting actual output before finalizing any task completion claim.

Evidence-based task verification in software engineering workflows is a protocol that links every completion claim to a specific, executed validation command and its actual output, eliminating unverified success assumptions.

Verify subagent work quality before final submission by applying structured quality guardrails that require the subagent to execute reproduction scripts and report the actual exit code and output to support success claims.

Mandatory verification works for general software testing and debugging workflows by applying a structured framework that requires fresh evidence and specific validation command execution before allowing any reliability or accuracy claims to be finalized.

The best way to eliminate the "it should work" failure mode in automated debugging is to require the execution of specific validation commands and the reporting of actual output to support every success claim before a task is completed.

You should not rely on agent self-reported task completion without fresh evidence when working on bug fixing, subagent delegation, or critical software workflows where reliability and accuracy require actual validation command output.