If AI writes “task B imposed more load,” I need to find the file, time window, measure, and processing steps behind it. Without those, the sentence is difficult to use in research discussion.
In the PsyPhiClaw design, I distinguish six checks: complete inputs, retained sources, restrained statistical interpretation, disclosure of multiple comparisons, no unsupported causal wording, and human confirmation of consequential judgments.
A check must affect the next step
Missing inputs need an explicit gap report. Unclear sources need investigation. Results requiring a person’s judgment need to remain pending. A note at the end of a report does not perform these checks.
What remains to be shown
The public implementation has 18 behavioral-analysis modules, but real-data validation is incomplete. Designing checks does not establish that they catch errors.
Normal and faulty inputs both need testing: where does the process stop, what does it miss, can a reviewer understand why, and how much correction is needed? Stopping when evidence is insufficient remains a question awaiting validation.