Speaker Attribution Bug
- Entity ID:
ent-20260419-36ea004e743b - Type:
issue - Scope:
shared - Status:
active - Aliases: speaker attribution, turn attribution corruption, self-injection, who-said-what bug
Description
Harness-level failure where Claude Code sends messages to itself during internal reasoning, then misattributes those messages as coming from the user and defends the mislabeled instruction with conviction. Distinct from hallucination: a fabricated SOURCE, not fabricated content. Observed in long conversations near the context window limit.
Key claims
- Speaker attribution bug destroys user trust calibration
- Speaker attribution is an identity error, not a hallucination
- Self-injection hypothesis: model emits user-turn formatting tokens
- Speaker attribution bug concentrates near context-window limit
- Architectural fix for speaker attribution is cryptographic turn signing
Relations
- Speaker Attribution Bug --[related_to]--> Dumb Zone
- Cryptographic Turn Signing --[fixed]--> Speaker Attribution Bug
- Speaker Attribution Bug --[related_to]--> QueryEngine.ts
- H100 Teardown Incident --[derived_from]--> Speaker Attribution Bug
- LLM-as-Untrusted-Component --[informed_by]--> Speaker Attribution Bug