Aglet

Verify an Adversarial Evaluation Slice

Verification keeps adversarial evaluation diagnostic. Review target, control, construction, variants, and expected response before reading the score. A passing slice reveals behavior under a defined condition and preserves limits on what the result means for decisions about related task slices.

Check whether the outcome improved

  1. Set adversarial checks

    Require target behavior, clean control, pressure condition, item provenance, expected response, evaluator view, severity, alternate response, and generalization rule. Define false-positive and ambiguous-case treatment for each tested slice and reviewer.

  2. Review controlled variants

    Apply checks to clean, mild, strong, conflicting, distracting, incomplete, and recovery cases. Inspect outputs and traces beside task contract. Record whether the control passed and the stressed case isolated the target weakness.

  3. Approve the slice scope

    State which behavior, pressure conditions, task family, and decisions the result supports. Record owner, reviewer, fixture, date, and retest trigger. Mark it inconclusive when control or evaluator evidence is missing.

What to carry forward

The verification record should show target, control, variants, evidence, severity, and scope. Approve only the susceptibility observed under the tested condition. Reopen it when task contract, pressure construction, context, evaluator, or target behavior changes. Record the clean control and claim boundary before using the slice to change agent decisions.

Technical background: Google DeepMind evaluation research.

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