Show the search, not just the answer — reasoning made auditable

The quiet interpretability advance of the cycle isn't a new probe into a model's internals. It's that a frontier result arrived with a trace of how it was found and a proof anyone can check.

Alongside its results, OpenAI published a reconstruction of how Astra searched for the arguments — the reasoning, not just the polished proof. Showing the search is a different transparency than showing the answer: it reveals the dead ends, the heuristics, where the key move came from, which is closer to understanding a model's reasoning than reading its verdict.

Auditability can stand in for understanding

And where the output formalises, you get something stronger still. Machine-checkable proofs turn AI reasoning into something auditable end to end — you no longer have to understand why the model believes a result, you can mechanically verify each step. That is a stronger guarantee than explanation, for the domains that admit it.

Complement, not replacement

This reframes interpretability's relationship to formal methods. Behavioural interpretability asks 'can we trust the reasoning'; formal verification answers 'we don't have to — we can check it.' The two are complementary: certificates handle the outputs that reduce to proofs, mechanistic interpretability remains necessary for everything that does not.

The frontier is enlarging the auditable region. The more of a model's work that can be emitted as a checkable certificate — proofs, typed programs, verified plans — the smaller the surface that must be trusted on faith. That is a real, if narrow, path to trusting frontier reasoning.

ExplainX — OpenAI Astra's ten math proofs explained → · arXiv — The agentic researcher: AI-assisted research in mathematics →