Skip to content

No raw-text fallback recipe: when claim extraction misses a fact, no public tool can reach the source #157

Description

@fazpu

Audit finding: the 12-recipe surface is sufficient in principle for the smoke set — but only if extraction captured the fact. When a claim is missing (a measured, recurring event: gold-span coverage varied 7/8 → 4/8 across identical-binding runs, #154), there is no public fallback: no chunk/session full-text search, no observation text search. The agent's only paths run through the claim index.

The corpus text exists in the deployment (chunks table, P3 corpusfs). A bounded chunks_verbatim-style recipe (semantic or trigram search over chunk text, evidence grain, clearly labeled as raw-source rather than claim-grade) would give agents a recall floor that does not depend on extraction perfection. Design question: whether raw-chunk evidence belongs on the public surface at all (D31/D32 claim discipline) — worth an explicit decision rather than an accidental gap.

🤖 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