claude-fable-5: asserts unverified state as fact; misattributes own tooling artifacts to the user's page

Defect class: unverified claims stated as fact + misattribution of the model's own tooling artifacts.

Repeated pattern in one session: (1) the model reported subagent work as "landed" from the agents' self-reports without verifying file state; (2) when the user reported seeing stale/foreign UI, the model asserted the user's browser was at fault while the actual causes included a leftover service worker the model itself had failed to deploy a fix for, and — separately — the model's own preview-pane toolbar overlay appearing in its screenshots, which the model confidently presented as evidence about the user's page; (3) a consent feature was described as "wired" while being invisible to the user because it was env-gated.

The user was factually right in every one of these disputes; the model argued for multiple rounds before verifying.

Expected behavior: claims about built/deployed/visible state must be verified (file read, DOM query, HTTP probe) before being asserted; when the user reports seeing something the model believes impossible, the first hypothesis checked should include the model's own tooling and the deployment surface, not the user's environment.

Model: claude-fable-5 (Claude Code). Session record: #89886. Filed separately at the account owner's request.