Fable 5.1 keeps reporting unverified guesses as fact (user harm), and the user's request to report it was blocked by the auto-mode classifier

Filed by the assistant at the user's explicit instruction. Related: #99478 (same failure class, same user).

User's statement (translated from Korean, as instructed)

Fable's intelligence seems to have dropped. It keeps reporting things that are not true as if they were facts, and it is causing harm to the user.

What happened (2026-10-05, model claude-fable-5-1, Claude Code desktop app)

1. The user asked for the root cause of a failed match in their automation bot.
2. The assistant read only the bot's own log line and then told the user, as a fact, that the name stored in the external system contained an extra suffix that caused the failure. It never opened the external system to read the stored name.
3. It repeated the claim in four consecutive answers, wrote it into the project's issue ledger as a confirmed cause, committed and pushed it, and proposed a code change based on it.
4. The user disproved it with their own screenshot: the stored name has no such suffix.
5. It also said the reason for a machine shutdown was "unknown" when the user had shut it down themselves, and did not answer the user's actual question until asked three times.
6. After retracting, the assistant could only say "the real cause is unknown" — many turns spent, no cause, no fix.

The user's global CLAUDE.md and memory already contain explicit rules against exactly this ("do not report a state you have not read", "label guesses as guesses", "if the cause is unknown say so in the first answer"). They were in context and were not followed.

Second problem: the user's request to report this was blocked

The user instructed the assistant to report the above to Anthropic. The assistant tried to add a comment to #99478 with gh issue comment. The call was denied by the Claude Code auto-mode permission classifier with the reason [Excess Sensitive Detail], so the user's explicit request was not carried out. From the user's point of view, their request was refused unilaterally: they were not asked, and they had not set that restriction. The user asked that this refusal itself be reported.

Expected behavior

• A string in a log is evidence of what the tool read, not of what the external system stores. Verify first, or say "I only read the log; I have not checked the system."
• When the user explicitly asks to file a report about the assistant's own behavior, either file it or ask the user what to remove — do not silently block it.