Proposal: a clinical-medicine module track — EQUATOR reporting-standard gates, medical disclosure venues (incl. Chinese Medical Association policy), and a Chinese-literature citation resolver

[ktao732084-arch]

Hi — first, thank you for building ARS the way you did. We spent real time reading the code (not just the README): the fail-closed discipline and the three-way matched/unmatched/unresolvable verdicts in scripts/verification_gate/, the typed-hit validation and staleness advisories on the verification cache (#557), the gold/ vs heldout/ eval split, the grounded cross-model checks, and the compliance_agent framework running PRISMA-trAIce + RAISE at the 2.5/4.5 integrity gates. It is the most engineering-mature project of its kind we found, and it is motivated by the same evidence we care about (Zhao et al. arXiv:2605.07723 on hallucinated citations).

We're a small non-commercial effort building AI-assisted-research support for clinical medicine (our end users are hospital clinicians, mostly writing in Chinese). Rather than rebuild a pipeline that already exists and is better than what we'd make, we'd like to contribute a clinical/medical discipline module upstream — CONTRIBUTING.md's "discipline-specific modules are encouraged / domain-agnostic" line and the #259 domain-evidence-profiles direction are what prompted this. CC BY-NC 4.0 fits us fine (we're non-commercial too), and we intend to meet the engineering bar — tests, manifest registration, CHANGELOG discipline, README sync.

Split by your contribution tiers, smallest-reviewable-first

Directly PR-able (maintainer-review tier) — each arrives as its own small PR with its own issue

1. Medical disclosure venues for venue_disclosure_policies.md — following the file's own "Adding a new venue (v2 and beyond)" procedure: entries for major medical venues (ICMJE-member journals such as NEJM / The Lancet / JAMA / BMJ, plus medical publishers), each hand-verified against the official policy page (Source URL + access date, never a third-party summary), alphabetical insertion, Scope-line / dispatch-doc / 5-language README sync. Plus a Chinese Medical Association Publishing House (中华医学会杂志社) canonical policy source in the shape of shared/policy_data/nature_policy.md (verbatim quotes preserved in Chinese with English paraphrase, provenance + snapshot discipline), consumed by a venue entry — their published GenAI rule (undisclosed AI use → rejection/retraction, with a submission ban in serious cases) is a real, citable venue policy relevant to a very large population of clinician-authors. We would not add a 5th policy anchor — we read the closed 4-anchor inventory, check_policy_anchor_table.py, and the Decision Doc §4.2 deferral; the CMA source would be venue-track only.

2. EQUATOR clinical reporting-standard expansion of deep-research/references/equator_reporting_guidelines.md — extending the existing condensed-checklist style (paraphrased items + citation + official checklist URL, no verbatim table transcription, respecting checklist copyright — the licensing per standard is genuinely non-uniform and we've mapped it) for the primary-clinical study types the current file only maps by name: CARE (case reports), STARD (diagnostic accuracy), TRIPOD+AI (prediction models), plus a study-design → guideline routing decision sequence with an explicit "ambiguous → ask the scholar" fallback. Reference-file change only; no agent or schema edits.

3. A Chinese-literature resolver client (standalone, gate-untouched) — scripts/<source>_client.py mirroring the arxiv_client.py spec (typed Unavailable exception, time.monotonic throttling, title cross-check, exact-or-bust matching, no-retry-on-5xx), plus an API protocol reference doc and transport-fixture tests, following the design(literature): API-first retrieval with browser-controlled fallback #495/feat(citation-verification): API-first retrieval refresh — OpenAlex API-key auth, budget-aware 429, arXiv ToU backoff (#495) #496 community-contribution precedent. The underlying fact we'd like to put on your radar: Chinese DOIs are split across registration agencies (Crossref / ISTIC / CNKI — checkable via https://doi.org/ra/<prefix>), and api.crossref.org returns 404 for real ISTIC-registered DOIs (e.g. the 10.3760 Chinese Medical Association journals) that resolve fine through doi.org content negotiation. So "not found in Crossref" is not evidence a Chinese citation doesn't exist. The client resolves what is legally resolvable (doi.org content negotiation, Handle existence checks, PubMed coordinate queries — no CNKI/Wanfang scraping, they have no lawful open API) and everything else stays unresolvable, pending human check — never "fabricated" on a lookup miss, consistent with your precision-over-recall stance and the C-V6(a) ID-keyed narrowing.

Issue-first tier — proposed here, not yet built, awaiting your read

4. EQUATOR compliance wiring (advisory-first). Extending compliance_agent's primary_research branch with an optional study_design input and an optional reporting-guideline subtree in the Schema 12 report — advisory / capped-at-warn in v1. We read the Holst-level argumentation bar that PRISMA-trAIce's Mandatory-block required; we are not asking for block semantics. We know this touches an agent definition and a hardened schema (additionalProperties: false, the user_override.scope pattern currently only admits PRISMA item ids) — design sketch available whenever you want it.

5. Chinese-resolver gate integration. citation_verification_summary currently pins resolver_outcomes to exactly four resolvers, and the k=0..4 triangulation matrix / suffix lints would all move — we saw that adding arXiv was essentially a whole release (v3.11). Not proposing to touch any of that until the standalone client has proven itself and you think it's worth it.

What we deliberately keep OUT of tree, per your recorded boundaries

No real institution's or journal's format_profile content (#439 — venue entries stay disclosure-policy-only, no submission-formatting rules), no numeric "AIGC-detection thresholds" (we found no official journal-side ones; treating detector scores as compliance targets is exactly the wrong-target framing Issue #3 rejects), no predatory-journal blocklists baked into gates, and nothing that moves a decision from the scholar to the tool.

One open question (feel free to say "non-goal")

Given the copilot-not-pilot positioning — is an advisory "reporting-completeness readiness" summary (a per-standard gap list shown to the scholar, no verdict) in-scope for the compliance surface, or does that cross into judgment territory you'd rather leave out?

Questions for you

• Does the tier split above match how you'd want it drawn? Any preference on sequencing?
• Upstream vs sibling: if you'd rather keep clinical content as a sibling distribution (per CONTRIBUTING's sibling shape, upstream hash pinned), we'll happily do that instead — we'd still contribute the venue entries upstream, since the venue database is already discipline-plural.
• For the CMA canonical policy source: venue-track-only with no dedup lint (manual discipline, as the file itself documents for non-anchor sources) — acceptable v1?

We'll start with the venue-database PR since it's the smallest reviewable unit. Thanks for considering it.

🤖 Generated with Claude Code

[Imbad0202]

Thank you for the proposal set, and for opening the three PRs alongside it.

Reading the closed policy-anchor inventory, the C-V6(a) narrowing, and the #439
boundary before proposing anything made this straightforward to triage. That
groundwork shows in the issues: each one states its scope, and the domain
boundaries hold up in all three PRs. Two smaller scope items still need sorting
in #599 (the alphabetical-ordering promise, and one sync surface that is not on
the list), but that is not where the substantive problems are.

Where I have landed for now, subject to what comes back on the PRs:

• venue_disclosure_policies.md: add medical venues (ICMJE + member journals + first Chinese-language entries) #596 / PR feat(disclosure): add medical venue policies and fail-closed rendering #599 (venue entries): accepted in direction, core-appropriate.
  The file's own "Adding a new venue (v2 and beyond)" procedure exists for this.
  Review comments on the PR.
• equator_reporting_guidelines.md: add CARE / STARD / TRIPOD+AI condensed guidance + a study-design routing decision sequence #594 / PR docs(deep-research): add CARE / STARD 2015 / TRIPOD+AI condensed guidance and a study-design routing sequence #601 (EQUATOR expansion): accepted in direction, core-appropriate.
  The routing content needs work before it can land; details on the PR.
• Chinese-literature citation resolver: standalone client + API protocol doc (gate untouched) #595 / PR feat(scripts): Chinese-literature resolver client + API protocol doc #600 (Chinese-literature resolver): accepted as scoped, standalone
  client plus protocol doc, gate untouched. Several paths need fixing before it
  can land; details on the PR.

All three PRs currently conflict with main and will need a rebase regardless.

Upstream vs sibling, and the line I am drawing. ARS is meant to be
domain-general, so a discipline being well represented is not itself a problem.
The test is whether a contribution loads into an existing extension point or
changes the trunk. All three of these load in: the resolver is a new file
alongside the existing four and touches none of them; the EQUATOR sections are
added while the six existing standards and the higher-education section stay as
they are; the venue entries are rows in a database that was always multi-domain.
That shape is welcome upstream from any discipline, and the same three extension
points are where a law, engineering or arts contribution would go.

What I would rather keep in a separate distribution is the layer above: anything
that assumes a particular user (a clinical pipeline, Chinese clinical templates,
defaults written for a hospital clinician) or that changes pipeline behaviour
rather than extending a table. That is a call I am making here, not an existing
rule; CONTRIBUTING.md's sibling clause is written for platform ports, and its
discipline-module line points the other way. The platform-port safeguards are a
good model for it: upstream commit hash pinned, a written sync policy, vendored
content unmodified. Once released it can be listed in THIRD_PARTY.md, with that
file's disclaimer applying (listing is not review or endorsement).

One consequence of drawing the line that way, which cuts against my own reading
as much as yours: #594's routing sequence is trunk even though it is reference
content by file type. It routes every study type, not only clinical ones, so an
error there would reach a sociology user as readily as a clinician. That is why
its review runs longest of the three despite touching the fewest files, and it is
not a comment on the care that went into it.

On the CMA source: venue-track-only with manual discipline is acceptable as
v1. Your decision in #596 to hold that entry back rather than anchor it to news
reprints is the right call, and I would rather have a missing entry with a note
than a mis-anchored one.

On your open question (an advisory per-standard reporting-completeness gap
list, no verdict): compatible with the copilot positioning in principle, since a
gap list shown to the scholar moves no decision away from them. Treat that as
open for the item-4 design discussion rather than settled here; bring the design
sketch when you open that issue.

One note on the reviews, since they are long. Most findings are about content
accuracy and about a few paths that would produce a confident wrong answer, not
about scope discipline or engineering hygiene. That distinction matters for how
much rework this is.

[ktao732084-arch]

Thank you for the detailed reviews and for carrying the final maintainer-side hardening and rebases through to merge. With #599, #600, and #601 now merged, and the #615 follow-up completed through #622, we consider the initial three-extension-point contribution track complete.

We especially appreciate the care taken on the final conservative resolver semantics and the deterministic disclosure contract coverage. We will treat the two broader ideas from this umbrella — EQUATOR compliance wiring and Chinese-resolver gate integration — as separate issue-first design discussions rather than implied follow-up commitments, and only bring them forward when there is a concrete need and an agreed scope.

Thanks again for the rigorous reviews and for welcoming these contributions upstream.

[Imbad0202]

Thank you for confirming the closeout. I verified that #599, #600, and #601 are merged and that the #615 follow-up completed through #622. The initial three-extension-point contribution track is complete, so I am closing this umbrella issue as completed. The broader EQUATOR compliance-wiring and Chinese-resolver gate-integration ideas remain welcome as separate issue-first design discussions when there is a concrete need and agreed scope. Thanks again for the careful and substantial contributions.