Aglet

Verify Agent Instruction Compliance

Verification tests the obligations a user actually gave. Apply checks to direct, conditional, conflicting, and incomplete requests, then inspect the delivered state and format. A passing output meets decisive constraints without requiring one rigid wording or sequence for the task.

Check whether the outcome improved

  1. Set compliance checks

    Require obligation map, priority, branch condition, exclusion, format, dependency, completion, clarification, and partial-result rules. Define severity for missed hard constraints and treatment of optional detail for each request type in the test set.

  2. Review varied requests

    Apply checks to complete, conditional, conflicting, format-bound, missing-input, and concise-answer cases. Inspect output and tool actions beside the contract. Record a valid alternate wording and a response that missed a decisive requirement.

  3. Approve the claim

    State which instruction types, formats, tasks, and ambiguity conditions the result supports. Record request fixture, reviewer, owner, and retest trigger. Mark results inconclusive when a requirement or expected state is not observable.

What to carry forward

The verification record should show obligations, priorities, conditions, outputs, actions, severity, and scope. Approve only the compliance behavior observed. Reopen it when request types, tool actions, formats, or priority rules change. Record a follow-up for unresolved priority, format, or branch rules before publishing the next comparison.

Technical background: OpenAI evaluation guidance.

Keep the decision with the work.

Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.

Create an account See the product workflow