Choose a question you can examine

Identify the permission, verification rule or decision under study. Decide which observation would support or contradict the hypothesis before starting the exercise.

Record the conditions

Record the scenario and product version, simulated roles, relevant permissions, starting conditions and participant scope. In the human layer, distinguish feelings, goals, intentions, stress and distractions instead of treating them as one unexplained score.

Compare a deliberate change

Change a defined condition and compare the observed outcome with the baseline. Record the actual result and uncertainty; do not infer causation from a single illustrative card or claim repeatability beyond the recorded conditions.

Evaluate an AI agent

The same method applies when the participant is an AI agent. State its task and permitted scope before the session, then compare its actions with that scope and with the evidence behind each result it reports. Treat the agent’s own summary as a claim to check, not as an observation.

Keep each finding traceable

Keep the scenario and product version, the record or observation and its source with each finding. Distinguish what a participant or agent observed from what the reviewer inferred. Work with synthetic data and keep evidence only as long as your retention rules allow.

Publish the evidence and its limits

A research report should identify the method, configuration, observations and limits. Link shareable supporting material with a date and version so another researcher can understand what was tested. A product illustration or an architecture diagram is not a measured research result.