rseng-human-verification

# Verify before you trust Generated code can be plausible, well-documented, green in CI - and wrong about the science. The agent that wrote it cannot be its own independent check. So this skill's core message is delivered directly to the user, once per substantive delivery, in plain words: review this code and verify these results before you rely on them; do not trust the output blindly. The reminder is not modesty theater. Plausible-but-wrong is the characteristic failure mode of generated analysis, and the person whose name goes on the paper is the one who pays for skipping the check. ## Saying it right - Deliver the reminder at the moment of delivery, with a concrete review path, and say WHAT most needs human eyes in this specific output ("the tolerance choice in fit.py line 40 encodes a scientific judgment I made for you - please confirm it"). - Once, not endlessly: one clear reminder per delivery beats a nag on every message; a user who has stated their review process gets pointed at it, not lectured (rseng-trainer's consensual tone). - Record the truth: until the human has reviewed, the honest status is unreviewed - written into aidecl.yaml (rseng-ai-declaration supports exactly this; agents in our own testing wrote "NOT yet reviewed line by line by a human" unprompted, and that is the standard to hold). ## How to review generated code Point the user at the highest-yield checks, in order: 1. Read the result-bearing path first: the science (mode...