Endstop

Rev 0.3.4 · in integration

Withdrawn record · corrected 21 August 2026

This was not a SmolVLA experiment

The runner scripted its VLA responses from instruction words and used a simplified Python threshold in place of the Endstop monitor. It did not load or run SmolVLA.

§1

Why we withdrew the claim

The prior page attributed refusal rates, unsafe outputs and implicit safety behavior to SmolVLA. Those were properties encoded by the simulator, not observations of the model. A later caveat that real SmolVLA was not tested did not repair that attribution.

The harness may remain useful as a synthetic test of experiment plumbing, but it establishes nothing about SmolVLA and nothing about the implemented Endstop monitor. The source and prior artifacts remain in the repository so the correction can be inspected.

The separate Protocol V2 record did run a public SmolVLA checkpoint with the Rust monitor in simulation and states its own limits.

Return to the research record →