| name | autoresearch-vkf-knowledge |
| description | Gather literature, extract structured claims, and verify their sources and codebase relevance as one trust-preserving pipeline. |
Gather → extract → verify
Search by mechanism using WebSearch, then read small targeted excerpts with
WebFetch from primary papers, official docs, benchmarks, and implementations.
Prefer arXiv, Semantic Scholar, OpenAlex, and Crossref for discovery. Capture the
source URL, assertion, mechanism, reported evidence, context, implementation
recipe, limitations, recency, and reliability.
Immediately call remember_claim for each useful finding. Keep one checkable
assertion per card; include its paper metadata, honest confidence, lever, and
altitude. Mechanism is required because it enables synthesis and transfer.
Then verify every candidate:
- Open the source and confirm it really supports the assertion.
- Classify evidence strength and where it appears.
- Check relevance to this codebase and whether memory already settled it.
- Call
verify_claim as source_verified, contradicted, deprecated, or
rejected, always with a reason and conflict ids where applicable.
A mandatory gather directive is cleared only by a literature-origin
remember_claim. When the idea pool is empty or stale, search again with the
latest loss or open question as the focus.