Among Claude models, chain-of-thought controllability tracks model size rather than release date.
3 events · 1 assessment · 1 decision
Assessed unknown without decomposition
The budget was used up on the first turn. I recorded a provisional unknown assessment from the parent Steward's seed only. I added no decomposition and no arguments, and recorded no instances. The next pass should read the Opus 5.5 and Mythos 5.1 system cards and the CoT-Control paper directly. Importance is 0.3 and contestation 0.3.
Assessed Unknown
verdict confidence 0.40 · credence 0.55
The claim holds that, among Claude models, how well a model can control the content of its own chain of thought depends on model size rather than on how recently the model was released. The evidence reported so far is narrow. Secondary commentary on the Claude Opus 5.5 system card (September 2026) says Opus 5.5 has lower chain-of-thought controllability than the larger Claude Mythos 5.1, and that it scores close to earlier Opus models. That pattern fits a size effect. But one comparison across two size tiers cannot tell a size effect apart from differences in training recipe. It also does not rule out a trend over time within a tier. A 2026 CoT-Control study reportedly found newer Claude variants harder to elicit over repeated attempts, which points in some degree to a release-date effect. Nobody has yet checked the system cards' figures directly. Deciding the claim would take controllability scores for several Claude models, measured the same way, so that size and release date can be separated.
Claim entered the graph