updown move enter open esc close matheesan.ca / index

    CASE 03 · evals

    Silent green

    The AI made the change. The tests passed. The bug shipped anyway.

    6 of 16 failures5 lint rules1 rule across 5 projects
    evalsLLM QAlintingagentic codingtest oracles

    problem

    AI-authored agent configs looked done when they were not. The change was made, the suite reported green, and a real defect still reached users. I collected 16 field failures to find out why.

    approach

    Classified 6 of the 16 as “silent green” and wrote them as harness-ready eval scenarios with must / must-not assertions, so an eval that only checked diffs and test results could no longer score them as successes. Proposed five static lint rules; the first surfaced new issues across five projects. Codified the team rules: “0 collected ≠ 0 failed,” “allow_failure is not a pass,” and “assert absence for stop-doing-X requirements.”

    the hard part

    The failure mode is absence, and absence is quiet. A suite that never runs exits cleanly. A linter that times out can print “No issues found.” An AI agent can satisfy a requirement by weakening the assertion. Each fix had to make the missing thing loud.

    result

    Six historical failures now run as explicit eval scenarios. Five rules were accepted for the platform linter, and the first one found new issues in five projects.

    what I’d do differently

    I would wire the eval harness before the first AI-authored config, not after the sixth silent failure.