Claude Fable 5

Thinking was recorded rather than refused, by prior registration

Alone among the models on this instrument, this one was registered in advance to have its thinking RECORDED rather than suppressed, and a nonzero count on it is not a gate failure. That was registered before the run precisely so that a result on this model could not later be explained away, or quietly discarded, on the grounds that it thought. The figures beside this are a measurement; the zeroes on the other models are a refusal holding, and the two are not the same kind of number.

Records reporting a nonzero count

1048 of 1358

Thinking tokens recorded

148,834

The plan served a different model on some calls, and those calls were refused

Some calls asking for this model came back answered by another one: same prompts, same flags, no error, and a plausible score. Nothing in the envelope announced it except the served model string the adapter now records on every record. A record served by another model is evidence about that other model, gathered under conditions chosen for this one, so it is invalid and never scored. It is not deleted and its tokens still count against the run's spend, because the call was made whoever answered it.

Calls served by another model

48 of 1358 on the tips replay, spanning 28 unit-runs across 10 distinct claim, task and arm unitsServed instead by Claude Opus 5, on 2026-08-31.

Skill cells withheld rather than thinned

2Each is published as withheld, with its reason and its surviving pair count, on the skill’s own page.

Prevented from

2026-08-31The served model is read back from the CLI’s own usage envelope and recorded on every record, and a record whose served model differs from the requested one is invalid at birth and invalid at load. Committed in 7ac6030 and 81253fc.

The finding in full, and the checks that catch it now.