// news · research-papers · alignment2026-08-09source: arXiv preprint

Reasoning Consistency Scanning audits whether chain-of-thought is valid, not just present

A framework for auditing chain-of-thought validity inside AI safety evaluations. The premise is that a safety evaluation which reads reasoning traces needs to know whether those traces are sound before treating them as evidence.

The gap it addresses is one most evaluation pipelines skip. A trace is treated as a window into the model's process, but a trace can be internally inconsistent, circular, or arrive at the right answer through invalid steps — and none of that shows up if you only check the conclusion.

Consistency scanning is the tractable version of the question. Full validity checking is undecidable in general; checking whether a trace contradicts itself is mechanical, and it catches a real class of failure.

It belongs alongside the monitorability work rather than competing with it. Monitorability asks whether a monitor can read the trace; this asks whether the trace deserves to be read.

See our analysis →

arXiv — Reasoning Consistency Scanning: a framework for auditing chain-of-thought validity in AI safety evaluations →