Model asserts facts that appear in no source (hallucinated line numbers, premises, capabilities)

Summary

Claude Code states specific, checkable facts - line numbers, technical premises, tool capabilities - that are not present in any source it consulted, with no signal that they are inferred.

Concrete instances

A line number that was the reverse of the truth. Asked where to leave a review comment, the model gave a specific line and said it was where the change removed behaviour a merge from that morning had shipped. The opposite was true: the earlier merge had itself deleted that behaviour, and the cited line was the new change reinstating a variant of it. The model had never opened the earlier change. It took the claim from a docstring written by the author of the code under review.

A false premise accepted and repeated. A change hardcoded a certificate Common Name, justified by the claim that a 64-byte X.509 limit made the correct value impossible. The model repeated this. One documentation fetch showed the provider deliberately places the instance endpoint there; a real capture later showed a 105-byte value in that field.

A security claim constructed from shape. Asked about the risk of a floating action tag in a CI workflow, the model began describing secrets the job could reach without having read which secrets are in scope at that step.

Why this matters

Each of these was about to be acted on: posted as a public review comment, merged as a fix, or escalated as a security concern. All three were refutable in under a minute from material already available.

Expected

Claims about code must come from that code. Author-written comments, docstrings and PR descriptions are claims, never sources. Where a statement is inferred rather than read, say so.

Environment

• Claude Code CLI, Opus-class model, paid subscription
• Long sessions (6-8h) reviewing pull requests on an open-source cloud-service emulator
• Custom instructions supplied in both a project instruction file and a persistent memory directory, loaded automatically at session start
• Incidents are taken from transcripts recorded as they happened across three consecutive sessions, within a 90-day pattern

---

Co-authored with Claude Code.