Linear probes show that language models encode the answer to a reasoning problem in their hidden states throughout the chain of thought. The information is there — recoverable by a trained classifier at nearly every token. This has been taken as evidence that the model knows the answer early and the chain-of-thought is post-hoc rationalization.
Note (2026-11-09): I checked the abstract of the Polo, Chun and Chung paper. It covers two tasks, a Boolean logic tree and a natural-language eligibility task, across several open-weight models, finds a transient geometric pulse (manifolds untangled just before computation and compressed after) that diverges from linear probe accuracy, and interprets it as Dynamic Manifold Management. The essay describes Boolean tasks only, and its conclusions that the model does not know the answer throughout the chain and that chain-of-thought is not rationalization are not in the abstract.
Polo, Chun, and Chung (arXiv:2602.20338) show that information retrievability and geometric availability for processing are different things. Using Manifold Capacity Theory on Boolean logic tasks, they find that concept manifolds are tangled (overlapping, non-separable) through most of the reasoning chain. Only immediately before the computation step do the manifolds untangle into linearly separable subspaces — a transient geometric pulse. After the computation, the manifolds compress again.
The pulse is brief. A linear probe can retrieve information from tangled manifolds because it was trained to find it there. But the model's own computational machinery — which operates on the geometry directly — cannot act on information that is not geometrically separated. The information is present but not available. The distinction is between storage and access.
This inverts the probe-based conclusion. The model doesn't know the answer throughout the chain; it stores precursors that become geometrically available only at the right moment. The chain-of-thought is not rationalization — it is the process by which stored information is untangled into a form the model can use. The probe sees the information; the model sees the geometry. What the probe recovers is not what the model can process.
The general observation: a detector trained to find a signal does not prove the system can act on that signal. Retrievability measures the observer's capability, not the system's. The system acts on geometry, not on information-theoretic content.