Beyond Crossref/OpenAlex: Cross-Domain Research Methods Survey (2026-08-01)

Beyond Crossref/OpenAlex: Cross-Domain Research Methods Survey

Date: 2026-08-01 (Stage 2) Scope: How legal, medical, historical, journalistic, and general/consumer research validate sources and claims without a DOI-keyed academic citation graph — the exact gap in Vox’s current TrustScorer (Crossref retraction + OpenAlex venue reputation), which only functions for scholarly-paper-shaped sources. Companion to deep-research-domain-agnosticism-audit-2026-08-01.md.

Vox’s current trust/novelty scoring pipeline treats every source as if it were a journal article: Crossref answers “has this been retracted?” and OpenAlex answers “is this a reputable venue (journal/repository/conference/preprint)?” Both depend on a DOI. The moment a query touches history, law, current events, or consumer products, both signals go silent — not because the source is untrustworthy, but because the infrastructure that scores it doesn’t exist in DOI-space. Below are eleven concrete, structurally adoptable techniques from five fields that solve exactly this problem, each with an implementation note.

1. Legal research: citation validity graphs + jurisdiction weighting

Technique 1 — Shepardizing/KeyCite-style validity signal. Legal researchers don’t just check if a case exists; they check its current precedential status by walking the graph of everything that has cited it since. Shepard’s (LexisNexis) and KeyCite (Westlaw) assign color-coded signals — green (“good law”), yellow (“distinguished/criticized”), red (“overruled/reversed”) — computed by classifying every subsequent citing document’s treatment of the cited case (BYU Law Library, Westlaw KeyCite guide). Adoptable structure: don’t just check if a source is flagged (retracted); check the citation-treatment of everything downstream of it — has anything superseded, criticized, or reversed the claim since. Data: CourtListener / Free Law Project REST API is free, no-auth, and exposes exactly this — 9M+ case law decisions with a citation-network API. For non-legal domains, no equivalent free citation-treatment API exists — needs an LLM heuristic classifying each citing/discussing source as affirming, contradicting, or superseding.

Technique 2 — Primary vs. secondary source hierarchy with jurisdiction-aware weighting. Legal research ranks statute/constitution/regulation (primary, binding) above a law review article or treatise (secondary, persuasive only), further weighted by jurisdictional scope — a controlling appellate decision in the same circuit outranks an out-of-circuit or lower-court opinion (Monmouth Univ. Library). Adoptable structure: source-type tier + scope-of-authority weighting, not tier alone. Data: no clean free API for jurisdiction inference; needs LLM classification of “what authority/scope does this source claim.”

2. Medical research: evidence-hierarchy tiers + non-journal retraction signals

Technique 3 — Tag every claim with an evidence-tier, not just a source-tier. Clinical guidelines score the study design backing a specific claim (systematic-review/meta-analysis > RCT > cohort > case-control > case report > expert opinion), then apply GRADE to rate confidence per outcome, downgrading for risk of bias/imprecision/inconsistency/indirectness/publication-bias and upgrading for large effect size (GRADE overview, PMC). This decouples “is the venue reputable” from “is the evidence type for this specific claim strong” — the same journal can carry a case report (weak) and a meta-analysis (strong). Adoptable structure: classify each extracted claim by evidence-type tier, independent of venue reputation, surfacing both scores separately. Data: no API; LLM heuristic reading study-design language (“randomized,” “case report,” “systematic review of N studies”).

Technique 4 — Retraction/safety signals outside the DOI graph. Most medically consequential corrections never touch Crossref: a drug recall, a trial terminated for safety, an FDA communication. ClinicalTrials.gov’s v2 API returns a trial’s current status (terminated/withdrawn/suspended) by NCT ID, and openFDA exposes drug/device recall and adverse-event data, both free/no-auth (openFDA, ClinicalTrials.gov). Adoptable structure: when a source cites a clinical trial or drug, cross-check trial/recall status as an independent negative-signal check, parallel to Crossref.

3. Historical research: corroboration-as-trust-signal + provenance chain

Technique 5 — N-source corroboration in place of a retraction database. There is no “Crossref” for a 200-year-old letter. Historians treat convergence across independent sources as the trust signal itself — a claim gains credibility when sources that couldn’t have copied from each other agree, and loses it when it rests on a single, uncorroborated account (UMass Dartmouth). This is the single most portable technique here because it needs no external API — it’s a property of the retrieved set itself. Adoptable structure: for any claim with no DOI/retraction path, count independent corroborating sources (with an independence check — same author/outlet/wire-service origin doesn’t count as a second source) and let corroboration count substitute for the trust score Crossref/OpenAlex would otherwise supply.

Technique 6 — Primary/secondary/tertiary tiering + provenance/chain-of-custody. Historians rank an original artifact above an interpretive work above a summary-of-summaries (UMN Crookston), separately tracking provenance — a transcribed/translated copy is a strictly weaker witness than the original even if both are “primary” in genre. Adoptable structure: tag sources on two independent axes — genre tier and copy-distance-from-original — since OpenAlex’s single venue-reputation axis conflates both.

4. Journalism: independent-source counting + outlet-level correction track record

Technique 7 — The two-source rule as a formalized corroboration threshold. Traditional journalism requires a claim (especially a sensitive one) to be confirmed by at least two independent sources before publication. Same structure as Technique 5 but with an explicit minimum-count gate: do not surface a claim as high-confidence unless it clears a corroboration-count threshold, rising for high-stakes claim types.

Technique 8 — Structured verdict methodology from fact-checking orgs. PolitiFact and Snopes publish transparent methodologies; a comparison study found the two orgs agreed on all but one of 749 matching claims after normalizing rating scales — independently-run verdict processes converging is itself evidence the process is sound (Ballotpedia). Adoptable structure: when a fact-checking org has already rated a specific claim, surface that verdict as a first-class trust signal — the closest non-academic analog to a citation. Data: no unified free API; needs targeted WebFetch against known fact-check domains for checkable (political/viral) claims.

Technique 9 — Outlet-level correction-track-record as a trust prior. Transparent, frequent correction of errors is a positive trust signal for an outlet; a pattern of uncorrected errors is negative — the news-domain equivalent of retraction, scoped to the publisher not the article. NewsGuard operationalizes this: trained journalists score 2,100+ sites 0–100 across nine criteria including “publishing transparent corrections” (NewsGuard). Data: commercial/licensed, not free-tier — flag as a paid-API option or replace with an LLM heuristic checking a domain’s public track record.

5. General/consumer research: domain-reputation heuristics without any citation graph

Technique 10 — E-E-A-T-style multi-factor scoring for sources with zero formal citation infrastructure. Google’s Search Quality Rater Guidelines formalize what to check when there’s no DOI, no peer review, often no byline: Experience (demonstrable first-hand use), Expertise (accurate terminology, structured depth), Authoritativeness (referenced by other independent, already-trusted sources), and Trustworthiness — stated as “the most important member of the family” (SEJ E-E-A-T guide). Adoptable structure: replace the single OpenAlex “venue type” score with a multi-factor LLM-scored rubric for any non-academic source. Data: no clean API — explicitly not machine-computed even at Google; an LLM-prompting heuristic over source text + metadata.

Technique 11 — Recency-sensitivity as an explicit claim property. Consumer/current-events research treats “how fast does this fact decay” as a property distinct from source quality — a 3-year-old product review about a discontinued model is stale regardless of outlet authority, whereas a historical primary source doesn’t decay at all. Vox’s current pipeline treats publication date as roughly constant-value metadata rather than a decay function tied to subject volatility. Data: no API; a query-time classification (“fast-decaying topic — prices, availability, current events — or slow-decaying — history, established facts?”) setting a recency-weight multiplier, reusing date metadata already collected.

Synthesis: top 5 additions for Vox’s trust scoring

Vox’s current scoring answers exactly one question — “is this a legitimate, unretracted piece of the academic literature” — and goes silent for anything without a DOI, which is most of history, law, journalism, and consumer research. Priority order:

Independent-source corroboration counting as the universal fallback trust signal (Techniques 5 & 7). Highest leverage: needs no external API and covers every domain Crossref/OpenAlex can’t reach. Count independent, non-derivative corroborating sources when there’s no DOI, with an LLM independence-check filtering out sources sharing a common wire-service/press-release origin.
Per-claim evidence-tier tagging, decoupled from source-tier (Technique 3, generalizing Technique 6). Score the evidence type behind each specific claim (primary/eyewitness vs. secondary analysis vs. tertiary summary vs. opinion), not just the venue.
Domain/outlet-level reputation and corrections-track-record scoring for non-academic sources (Techniques 9 & 10). A cacheable per-domain LLM-rubric reputation score as the non-academic analog to OpenAlex venue reputation.
Domain-specific validity/status APIs wired in alongside Crossref, not instead of it (Techniques 1 & 4 — CourtListener for legal, ClinicalTrials.gov/openFDA for medical). Both free, no-auth REST APIs slotting into the existing “check if this record has been superseded/invalidated” pattern Crossref already implements, keyed to a different identifier (case citation, NCT number, drug name) than a DOI.
A recency-decay weight keyed to claim/topic volatility (Technique 11), cheap to add via query-time LLM topic-volatility classification reusing date metadata already collected.