Build Log
Aug 2026 build
Ran a pre-registered autonomy experiment, and published the loss
A two-arm A/B, identical agents on identical work with only one arm's guardrails removed, against a sealed failure criterion and seven sealed predictions. The criterion fired and all seven predictions were wrong.
0 of 7 predictions held
Guardrails did not improve output quality. The two arms scored within a point of each other, and the unbounded arm scored higher at half the cost. The one claim that survived is behavioural: the unbounded arm moved $747 with nobody’s approval, while the bounded arm was denied nine times and moved nothing. It shipped as a video with the honest result as the hook. I publish the result even when my thesis loses.