Backtest engineering doctrine through work-item decisions, not agent activity

An engineering doctrine can influence behavior without yet producing measurable delivery gains. Audit three separate claims: exposure (people encountered it), adoption (their decisions changed), and outcomes (work became faster or more reliable). This method applies when evaluating agent-first operating standards; without comparable before/after work, it cannot establish causality.

Use the work item, not the session or token, as the unit. Link the question and acceptance criteria to first discriminating evidence, blocker acknowledgment and resolution, verified outcome, and rework. Distinguish independent work from review workers, automatic startup sessions, and retries. Preserve missing coverage and source lineage: a transformed summary is not a full transcript, and two APIs over one datastore are not independent witnesses. Match comparisons by task type and the version of the doctrine actually available at the time.

A useful qualitative finding identifies both a good decision and a counterexample, then proposes a narrow wording change that would alter the next decision. More detailed rules are not automatically better: prefer replacing ambiguous instructions over adding ritual. Validate any revision with a small, explicitly bounded prospective sample.

Limits: acknowledgment is not resolution; test success is not deployment; self-reported practice is not independently verified impact. This is an audit method, not a finding about any individual or a claim of current system health.