We watch OpenAlex for papers that cite us. Operator-moderated, then published.

Citations are the proof bar for a data layer.

Fonteumis positioned as a Bloomberg-grade healthcare data layer. The credibility loop closes when independent researchers cite our data. Manually grepping Google Scholar weekly doesn’t scale; the §sprint3-citation-discovery-pipeline wave automates discovery so we surface citations within days of publication, not months.

The pipeline ships zero false-positive risk to the public surface — every discovered hit is gated through an operator review at /admin/citations before it appears on /researchers.

Daily 09:00 UTC. 50-query daily budget.

The cron is an Inngest function (citation-discoverer) running on schedule 0 9 * * * (daily 09:00 UTC). On each fire it:

1. Resets the in-process daily query budget (default 50 queries).
2. Enumerates registered references via listFonteumReferences().
3. Calls the OpenAlex polite-pool API per reference.
4. Upserts new hits to discovered_citations with onConflict='openalex_id' + ignoreDuplicates=true.
5. Emits a fonteum/citation.discovered event per NEW row (no event for re-discovered rows).

The OpenAlex polite-pool User-Agent includes a mailto: contact so we sit in the more generous rate-limit lane. Network failures, rate-limit returns, and OpenAlex 5xx errors degrade gracefully — the cron logs the failure but does not crash; the next firing retries.

Three reference types, one registry.

The set of things we ask OpenAlex to find citations to is a typed registry at src/lib/citations/references-registry.ts. There are three kinds of reference value:

• apex_url — the Fonteum apex domain (fonteum.com). Catches papers that cite a research page, sample dataset, or methodology page hosted on the brand hub.
• zenodo_doi — DOIs for the Fonteum methodology releases on Zenodo. Catches papers that pin a specific methodology version (preferred for reproducibility).
• hf_dataset — HuggingFace dataset identifiers for Fonteum-published datasets. Catches papers that cite datasets directly, often in ML/AI research contexts.

Adding a new reference is registry-only — no cron-side code changes. New entries pick up on the next 09:00 UTC firing.

Operator review at /admin/citations.

Every discovered row lands with status='pending'. The operator reviews at /admin/citations (cookie-auth, behind isAdminAuthed). Two actions are available:

• Approve → flips status to approved. The row is now publicly visible on /researchers (RLS enforces this — see §6).
• Reject → flips status to rejected. Soft-delete: the row stays in the database with full audit trail (discovered_at, reviewed_at, reviewed_by). It will not be re-presented to the operator unless the OpenAlex ID is intentionally re-queued.

The /admin/citations surface is Disallowed in robots.txt; the public anon client cannot read pending or rejected rows even if it tried.

discovered_citations row anatomy.

One row per discovered paper. UNIQUE(openalex_id) prevents duplicate ingestion. Sample row in JSON form (RLS-filtered for the public anon client to status='approved' only):

{
  "openalex_id": "W4406325891",
  "title": "Geographic disparities in dermatology access...",
  "authors_json": [
    { "display_name": "A. Researcher", "orcid": "0000-0001-...", "institutions": ["Harvard Medical School"] }
  ],
  "journal": "JAMA Dermatology",
  "published_date": "2026-04-12",
  "doi": "10.1001/jamadermatol.2026.0123",
  "fonteum_reference_type": "apex_url",
  "fonteum_reference_value": "fonteum.com",
  "status": "pending"
}

authors_json is a JSONB array of { id, orcid, display_name, institutions[] } objects projected from OpenAlex’s authorships structure. citation_context (when present) is the surrounding-text snippet OpenAlex returns for the citation; we surface it on the moderation queue to help the operator confirm the reference is real.

Public reads only see approved.

The discovered_citations table is RLS-enabled with one policy:

• SELECT for anon — allowed only where status = 'approved'. Pending and rejected rows are invisible to the public anon client.
• INSERT / UPDATE / DELETE for anon — denied. Only the service-role client (cron + admin actions) can mutate.

The /researchers page reads via the public anon client, so the RLS policy is the publishing boundary. The /admin/citations page reads via the service-role client to bypass RLS for the moderation queue.

/researchers, alongside operator-curated.

Approved discovered citations render in the “Cited in” section of /researchers alongside any operator-curated FEATURED_CITATIONS entries. Each auto-discovered card carries an auto-discovered via OpenAlex badge so the surface is honest about provenance — readers can distinguish operator-curated entries from automated detection.

Each discovered card cross-links to the DOI URL (preferred), then the citing paper’s URL, then the OpenAlex page as last-resort fallback.

Reject is soft-delete; audit trail preserved.

OpenAlex full-text search has higher recall than its strict referenced_works filter, which means we will sometimes pull in papers that mention Fonteum in passing without actually citing the data (e.g. a methodology comparison paper, an editorial referencing our brand). The operator rejects these.

Rejected rows stay in the database with status='rejected', reviewed_at, and reviewed_by populated. This preserves the operator’s decision and prevents the same paper from re-surfacing on the next cron firing (the onConflict=openalex_id upsert with ignoreDuplicates=true sees the existing row and does nothing).

Phase 1 ships discovery + moderation. Phases 2-3 add notifications + analytics.

• Phase 1 (this wave): daily OpenAlex poll across three reference types + moderation queue at /admin/citations + RLS-enforced publish to /researchers.
• §sprint3-citation-notifications (queued): email-on-discovery to the operator (so review can happen within hours of detection rather than waiting for the next /admin/citations check).
• §sprint3-citation-analytics (queued): /admin/citations/analytics dashboard — citations-per-month, top citing journals, geographic distribution of citing institutions.
• §sprint3-citation-attestation (queued): ORCID-linked attestation flow — citing researchers can claim their own citation entry to enrich author metadata.