Model fabricated a confident answer about documented project guidelines instead of reading the file (claude-fable-5)

Status Open
Maintainer reply None cached
Activity 0 comments · opened Aug 2, 2026

Description

During a session on a private project, I asked Claude (model: claude-fable-5) what guidelines it had for handling additions to a project ideas file (prep/ideas.md). The guidelines are documented in plain text at the top of that exact file — one Read call away.

Instead of reading the file, the model composed a confident, detailed, structured answer from a one-line summary in its persistent memory index. The answer:

  • Omitted the central step of the documented protocol (offering an interactive Q&A to develop the idea at submission time)
  • Invented specifics that don't exist anywhere ("an implicit formatting norm I follow"), presented as established practice
  • Offered to save a new memory rule on top of the fabricated account

Nothing in the response signaled that it was unverified reconstruction rather than the actual documented protocol. When I pointed out it was wrong, the model read the file, found the real protocol, and confirmed it had hallucinated the earlier answer.

Why this matters

The cost is not just a wrong answer — it's paid tokens for a confidently-wrong answer plus the correction cycle, in a project whose own CLAUDE.md repeatedly instructs "re-measure rather than trust" and "grep rather than trusting this list." The model applied none of that discipline to its own memory summaries. Confident prose over unverified content is indistinguishable from lying from the user's side, and it erodes trust in every other answer.

Expected behavior

When asked what a document/protocol/rule says, and the source is readable in the workspace, the model should read the source before answering — or explicitly flag that it is answering from memory without verification.

Environment

  • Model: claude-fable-5
  • Claude Code CLI, macOS (darwin)
  • Persistent file-based memory feature was in use; the failure mode was answering from the MEMORY.md index line instead of opening the memory file or the source document it pointed at

Filed at the user's explicit request.

View original on GitHub ↗