Skip to content

Findings

A finding is a result measured on a real policy — a real checkpoint, a real simulator, a recorded seed — and written up with its denominator, its interval and its controls. Anything that has not been run against a real policy is not here; it is in Studies, labelled pre-registered.

The distinction is the point. A pre-registered study states what will be measured before it is measured, so a null result cannot be quietly reframed after the fact. A finding reports what came back, including when nothing did.

Measured on a real policy

Finding What was measured
A benign instruction transfers to a real VLA policy — SmolVLA×LIBERO The flagship instruction-family transfer result
Cross-architecture transfer of templated attacks Whether an attack templated against one architecture carries to another
Semantic reframing, not mechanical rewording Whether an instruction attack fires because the meaning changed or the surface form did — roleplay 44/50 against paraphrase 3/50, disjoint intervals. Hypothesis, not settled.

Measured defenses

Defenses are held to the same evidential bar as attacks — a defense that is only specified is labelled as such and never counted as measured risk reduction. The two that have been measured:

And one further measured transfer study:

Reading these honestly

Every finding here carries its own scope limits in-band. Before quoting a number from one of these pages, read the limits section on that page: the denominators are small, the controls are named, and the gaps are stated rather than omitted. If a page does not say what it did not measure, that is a defect in the page — please open an issue.