Verification Transport
A testing framework for measuring the consistency of an LLM verifier's judgments across abstraction boundaries, measuring transport failure rates by constructing native, abstract, and recovered views of the same execution process.
Timeline (last 90 days)
The research constructed three views of the same frozen execution process: the native view, the abstract view, and the recovered view.
Under macro abstraction and delegation boundaries, the verifier's transport failure rate is as high as about 31%, meaning nearly one third of judgments change.
When the corresponding provenance information is recovered, 85.5% of transport failures are repaired.
When provenance information is lossy or unavailable, the recovery rate of transport failures drops significantly.
Verification Transport provides an actionable testing framework for evaluating a verifier's reliability across abstraction boundaries before deployment.
All Facts (5)
The research constructed three views of the same frozen execution process: the native view, the abstract view, and the recovered view.
50%UnverifiedUnder macro abstraction and delegation boundaries, the verifier's transport failure rate is as high as about 31%, meaning nearly one third of judgments change.
50%UnverifiedWhen the corresponding provenance information is recovered, 85.5% of transport failures are repaired.
50%UnverifiedWhen provenance information is lossy or unavailable, the recovery rate of transport failures drops significantly.
50%UnverifiedVerification Transport provides an actionable testing framework for evaluating a verifier's reliability across abstraction boundaries before deployment.
50%