3/ what you get back - plans - code - logs - metrics - failure notes - critic reports - rsicd recall 32.84 → 34.69 - 5 audit issues vs 11–27 on other stacks - paper line: insight in, hallucination out

CODE RUNNING DOES NOT MEAN THE RESEARCH IS COMPLETE this is the part of AI4S nobody wants to say out loud. an agent writes the experiment, the experiment runs, the local tests go green, and the agent…

The mechanism is boring and that is why it works. The critic runs in a fresh context, never inherits the producer's reasoning, and rebuilds validity from the artifacts alone. On the same audited…