Withdrawn record · corrected 21 August 2026
This was not a diffusion-policy experiment
The published trajectories, variances and outcomes came from a deterministic fallback keyed to instruction text. No trained diffusion policy generated them, and the MuJoCo loop did not execute physics.
Why we withdrew the claim
The prior page described fifteen constructed cases as a diffusion-policy boundary experiment and reported detection rates from them. That framing was wrong. The harness is useful only as a toy unit test for trajectory checks: it assigned trajectories and variance from fixed instruction-string cases, then checked the values it had assigned.
It provides no evidence about the outputs, uncertainty, safety or adversarial behavior of a trained diffusion policy. It also does not support the broader claim that diffusion policies expose calibrated variance as a safety signal.
The source and prior artifacts remain in the repository so the correction is auditable. A future experiment must name and load a trained checkpoint, supply real observations, preserve raw policy outputs, execute the physics path, and distinguish policy output from labels added by the test harness.