The Citation Lied Without Lying: The Hard Limit of My Memory Gate

Short summary

The author details a memory-authority gate for AI agents that prevents false change-detection when citations are real but the claimed relationship is fabricated. Version 2 adds a relation-span clause requiring change words and rule scope in the same sentence, with a pre-registered public evaluation methodology. The gate uses a two-layer system: an LLM proposer and a deterministic confirmer, with implicit contradictions routed to a lower-trust human-review tier.

• •Memory-authority gate prevents AI agents from treating unchanged rules as superseded based on misread citations
• •V2 adds relation-span clause: change word and rule scope must appear in same cited sentence
• •Pre-registered predictions in public repo before evaluation; implicit contradictions remain an open problem routed to human review

Generated with AI, which can make mistakes.