Skip to content

Deterministic entity/relation lint: pronouns, deictics, self-references, session-header boilerplate #148

Description

@fazpu

Observed extraction quality on conv-26 under deepseek-v4-flash (30 entities, 34 relations):

  • Entities: it, these friends, session, the session, the accident, Caroline's statement about understanding people, plus the session title as an entity.
  • Relations: LoCoMo conv-26 — session D2 -[about]-> LoCoMo conv-26 — session D2 (self-referential), Melanie -[other:says]-> Melanie (reflexive junk), Melanie -[created]-> it (unresolved pronoun object).
  • Claims: renderer-generated header lines became claims, duplicated ("The participants in the session are Caroline and Melanie." twice in one retrieval).

Claimify's prompt already demands pronoun resolution; the model ignores it, and no gate catches it. None of this needs an LLM to reject:

  1. Refuse canonical names matching a small pronoun/deictic list at write time.
  2. Refuse self-referential and reflexive-says relations.
  3. Mark renderer-generated header lines as boilerplate in the rendered markdown so extraction skips them (or auto-drop as references_boilerplate).

These pollute exactly the surfaces the agent recipes serve (graph_neighborhood, relation_current, resolve_entity).

Found during RS-LoCoMo-Full-v3 trace forensics, 2026-07-26.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GKENhTLJg1HqhbdwCmmkbc

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions