Handbook

Task `doc_claim_extraction` v1

Extract candidate documentation claims (checkable statements of capability, direction, positioning, or process guidance) from a Markdown document. Every emitted claim is status: candidate — the LLM classifies, humans…

Updated

Task doc_claim_extraction v1

Summary

Extract candidate documentation claims (checkable statements of capability, direction, positioning, or process guidance) from a Markdown document. Every emitted claim is status: candidate — the LLM classifies, humans decide truth.

Inputs

Field Type Required Notes
markdown string yes Full Markdown content of the seed or page.
source_path string no Provenance hint recorded alongside claims.
temperature number no Default 0.05.
timeout_sec integer no Transport timeout override.

Output

JSON object with a claims list. Each claim: claim_text, claim_type (capability_fact | architecture_direction | positioning | process_guidance | roadmap_direction | fact), maturity_signal (vision | defined | demonstrated | unknown), section, confidence (0..1), and forced status: "candidate".

Invariants

  • Output claims are never marked canonical; status is always candidate.
  • Unknown claim_type / maturity_signal values are coerced to safe defaults (fact / unknown).
  • Claims without claim_text fail with SchemaFailure.

Failure modes

  • SchemaFailure — empty markdown, non-list claims, or claim without text.
  • ParseFailure — invalid assistant JSON (one repair attempt).
  • TransportFailure — model/provider errors.

Changelog

  • v1 — Initial doc-hydration claim extraction task (Hydration v2, Phase 5).