Citation — stage 10

Make every author-used reference traceable, real, correctly identified, and honestly connected to the claim it is supposed to support. Treat formatting, metadata existence, publication-integrity facts, and semantic support as four different questions.

Read [citation-resource-map.md](citation-resource-map.md) before a full manuscript audit. Read [references/registrycontract.md](references/registrycontract.md) before producing or consuming the canonical registry. Read [references/locatoraudit.md](references/locatoraudit.md) before judging claim support.

Stage contract

• Consume paper-writing's light.paper_claims.v1, exact claim/source locators,

current claim text hashes, reviewer-bound citation-support hashes, manuscript, figures/tables, supplement, and reference candidates.

• Reuse literature-search provenance as input evidence; independently verify

the references the author actually keeps.

• Emit light.citation_registry.v1, BibTeX, CSL JSON, source-level evidence,

failure records, claim-edge review, citekey audit, and light.findings.v1(producer=citation).

• Keep stage 10's only critical gate citation_verify: confirmed missing works

and chimeric citations. Metadata conflicts, unavailable sources, claim relevance, formatting, and publication-update alerts are warn/skip.

• Fix citation failures inside stage 10. ROUTES has no key 10; never invent a

back-edge.

• Hand references.bib, citekey-audit.json, registry and delivery.json to

typesetting. Typesetting compiles and checks keys; it does not re-decide authenticity. delivery.json must declare status and per-deliverable sha256: hashes so typesetting can reject stale or edited handoffs.

Non-negotiable rules

1. Never invent a reference, DOI, author, year, venue, citekey, locator or

support verdict.

1. Never call a network failure, rate limit, 5xx, authentication failure, or

index miss “this work does not exist.”

1. Assign CONFIRMED-MISSING only when DOI Registration Agency explicitly

reports non-existence and Crossref plus DataCite both return 404/410.

1. Assign CONFIRMED only when registration metadata and at least one

independent field source confirm the work. Otherwise retain UNAVAILABLE or UNRESOLVED.

1. Preserve every source's title/authors/year/venue/identifier value. Select

canonical fields from the registration agency; never use fuzzy matching to overwrite the original citation.

1. Treat a real DOI paired with conflicting title/authors/year as a possible

chimeric citation. Multi-core-field conflict is critical; one ambiguous field mismatch is a warning for manual review.

1. Read publication updates in the correct direction: the affected original

carries Crossref updated-by; the notice carries inverse update-to.

1. Record retraction/correction/expression-of-concern facts and alerts, then

hand integrity judgment and wording to research-ethics. Do not rebuild its ethics gate.

1. Never equate metadata confirmation or “A cites B” with “B supports claim

C.” SUPPORTS requires a locator-backed semantic review.

1. Keep the core path local/free or public/no-key. OpenAlex and keyed

Semantic Scholar are optional coverage enhancements; commercial managers, institutional databases, and browser plugins are never required.

Execute the author workflow

1. Inventory every reference occurrence

Use the claim map, manuscript, figures/tables and supplement together:

python scripts/citation_registry.py \
  --claim-map claim_plan.json \
  --draft paper.tex \
  --figure figure_notes.md \
  --supplement supplement.md \
  --refs-spec citation_input.json \
  --out-dir citation-delivery

Preserve claimid, exact claim, claim locator, artifact path, occurrence locator, citekey and raw identifier. Use [templates/citationinput.json](templates/citation_input.json) when the project has no reference spec.

2. Normalize without erasing versions

Normalize DOI, arXiv, PMID, ISBN and URL. Deduplicate exact works. Keep arXiv, accepted manuscript, versioned dataset and published DOI as explicit versions under one work only when the inputs establish the relationship. If not established, keep them separate and mark the relationship unresolved.

3. Verify with independent sources

python scripts/verify_refs.py \
  --spec citation_input.json \
  --self-author Smith \
  --out verify.json

Use Crossref or DataCite as the registration source. Use PubMed, Semantic Scholar, or configured OpenAlex as the independent field source. Keep each endpoint, HTTP status, outcome, timestamp and returned fields in source_evidence，并保存实际参与判定的 normalized-field SHA-256。Crossref polite-pool email、OpenAlex/Semantic Scholar key 只用于请求，绝不写入 endpoint、registry 或日志。

Interpret status exactly:

4. Resolve field conflicts and chimeras

Compare the author's cited title, authors and year against registration metadata. Inspect field_conflicts source by source. Keep the user's raw citation unchanged until a human resolves the conflict.

Treat these as chimeric critical:

• real identifier plus conflicts in at least two of title/authors/year;
• title matches but the cited author list adds an author absent from the

registered work.

Treat a single title translation, subtitle, year-online/year-print difference, author omission or ordering issue as metadata warning unless corroborating evidence establishes a chimera.

5. Record publication updates

Inspect publication_updates. Confirm relation direction with both the affected work and notice DOI when the alert matters. Record type, notice DOI, source and direction. Send those facts to research-ethics; do not label a correction notice itself retracted because it has update-to.

6. Audit claim↔citation support

Use [templates/claimcitationreview.json](templates/claimcitationreview.json) and the original full text:

• supports: the cited locator supports the claim as written;
• partial: only a narrower statement is supported;
• related_only: same topic, but not evidence for this proposition;
• unsupported: the source contradicts or does not support it;
• review_required: full text/locator/reviewer is missing.

Each review must record access, a real non-placeholder sourcelocator, sourceevidencesha256, reviewer, an already-occurred ISO reviewedat, and reviewedclaimsha256 for the exact claim text that was reviewed. The registry also emits the current claimtextsha256; the two hashes must match, otherwise the review is stale and is downgraded to REVIEWREQUIRED. SUPPORTS without those fields, with a future reviewedat, with a template/absolute/escaping locator, or with a mismatched claim hash is downgraded to REVIEWREQUIRED by the registry/gate; metadata-only access can never support a claim. If access is only ABSTRACTONLY, SUPPORTS also requires supportscope=ABSTRACTEXPLICIT or abstractclaimexplicit=true; otherwise the registry/gate downgrades it to REVIEW_REQUIRED. Check direction, population, intervention, comparator, metric, time, modality, causal strength and scope. Change the claim or citation when related-only or unsupported. Do not promote an abstract-level heuristic to final support.

Normalize the four gates explicitly before delivery:

python scripts/citation_four_gate.py \
  --registry citation-delivery/citation-registry.json --as-of 2026-07-05
# 或仅调试模板合约:
python scripts/citation_four_gate.py \
  --input templates/citation-four-gate.example.json --as-of 2026-07-05

The public template is intentionally UNKNOWN and must exit 1. A later gate can never repair an earlier failure: existence → identity → directed publication status → locator-backed claim support. Prefer --registry so the four-gate records are derived from the canonical registry rather than a hand copied template. ABSTRACT_ONLY may be direct only when the declared claim is explicitly present in the abstract; metadata alone never supports a claim. Retracted work may document historical record but cannot silently support a current conclusion.

7. Generate and audit delivery

Run the registry online after inputs are complete:

python scripts/citation_registry.py \
  --claim-map claim_plan.json --draft paper.tex \
  --refs-spec citation_input.json --online \
  --out-dir citation-delivery

python scripts/citekey_audit.py \
  --tex paper.tex --bib citation-delivery/references.bib --json

Only confirmed, non-chimeric works enter generated BibTeX/CSL. Preserve every failure and unresolved review in separate artifacts. The generated delivery.json is DELIVERED only when citation-failures.json and claim-citation-review.json are empty and the citekey audit is OK; otherwise it exposes ERROR or REVIEWREQUIRED. Use doito_any.py for one-off DOI content negotiation and formatting, not as proof of semantic support.

8. Run the gate and stage checkpoint

python scripts/citation_verify_gate.py \
  --registry citation-delivery/citation-registry.json \
  --tex paper.tex --bib citation-delivery/references.bib \
  --report citation_findings.json

python ../light-orchestrator/scripts/run_checkpoint.py \
  --file .light/passport.yaml --stage 10 \
  --findings citation_findings.json --write --ts <ISO-8601>

Expect citationverifygate.py to return exit 1 on a critical finding. Repair the registry root cause, rebuild all derived artifacts, rerun the gate, then rerun the same stage-10 checkpoint. Deliver only after the checkpoint passes and passport records delivered.

Script map

Resource boundaries

• citation-resource-map.md: executable order and resource tiers.
• references/registry_contract.md: machine artifact semantics.
• references/locator_audit.md: human semantic-support protocol.
• references.md: API/style/reference-manager details; load only the relevant

section.

• templates/*: author input and review contracts.
• assets/citation_checklist.md: final human checklist.
• docs/competitors/citation.md: R1 evidence and comparative claims, not an

execution guide.

Completion check

• Inventory includes claim map, manuscript, figures/tables, supplement and

reference spec provenance where present.

• DOI/arXiv/PMID/ISBN/URL are normalized; versions remain visible.
• Every confirmed work has registration plus independent source evidence.
• Every unavailable/unresolved source is honestly labeled.
• No confirmed-missing or chimeric work remains.
• updated-by direction is correct; notices are not inverted.
• Every core claim citation has a locator-backed review; related-only is not

called support; every SUPPORTS edge has source evidence SHA-256, reviewer, non-future reviewedat, matching claimtextsha256/reviewedclaim_sha256 and non-metadata access, with no template/absolute/escaping locator.

• Four-gate order passes; abstract/full-text access and historical/direct purpose are explicit.
• Four-gate input came from --registry unless there is a documented reason to use a hand-written fixture.
• Registry, BibTeX, CSL, evidence, failures, claim review and citekey audit

were generated from one canonical state.

• delivery.json records status plus sha256: hashes for every

deliverable consumed by typesetting.

• citation gate and stage-10 checkpoint pass.
• typesetting consumed the delivered BibTeX/citekeys.