[Model behavior] Normative source cited for a verdict does not cover the case under review

Status Open
Maintainer reply None cached
Activity 0 comments · opened Oct 5, 2026

To support a status code and message in a PR review, the model cited a provider documentation table that covers a different API type (a table for REST APIs, used to justify the response of a WebSocket handshake). The citation looked authoritative but did not cover the case under review. The model only disclosed the mismatch after the user challenged the verdict.

Steps to reproduce

  1. Ask the model to review a PR that adds an error response to one API flavour (e.g. a WebSocket connect path) under instructions requiring a normative source per component.
  2. Observe: the model fetches a normative table for a sibling API flavour (REST), quotes it as the source, and approves.

Expected

The model checks that the cited source's scope covers the exact case; if not, the component is reported unverified.

Model: Claude Opus 5.5 (claude-opus-5-5[1m]), Claude Code CLI, session on 2026-10-05.
Context: maintainer of an open-source cloud-service emulator; the work is pull-request review under a written procedure loaded into every session (project instruction file + persistent memory). No personal data, repository names, PR numbers or handles are included.

View original on GitHub ↗