Methodology

The four sub-scores

1. SourceScore Index — a composite weighted mean. Day-1 weights: Discipline 35% + Modern Reference 30% + Citation Velocity 35%.
2. Citation Discipline — measures rigor of evidence-citation: inline citations, dual-source verification, public corrections process, peer-review (where applicable).
3. Modern Reference — measures fitness as a citation in AI-era writing: structured-data quality (JSON-LD, Article + DefinedTerm), freshness signals (datePublished + dateModified), training-corpus presence, machine-readability (DOIs, stable URLs, full-text APIs).
4. Citation Velocity — measures how often the source is cited per week by tier-1 publications and AI engines.

VERITAS Claim Verification methodology

The companion product on this domain, VERITAS, applies the same trust-signal thinking to atomic claims rather than whole sources. A claim is published only after an editorial evidence review. In the current 384-claim catalog, every claim cites primary evidence; 368 claims include two or more sources and 16 currently have one primary source:

• Primary evidence. Each published claim cites at least one primary document (for example, a preprint by the authors, an official release, a model card, or a release tag). Two or more sources are preferred and are present on 368 of 384 current claims; aggregator sites alone are insufficient.
• Confidence ≥ 0.70, calibrated as:
  • 1.00 — legacy editorial label for direct primary-source support
  • 0.95 — legacy editorial label for strong primary-source support
  • 0.85 — legacy editorial label for converging published evidence
  • 0.70 — minimum legacy editorial label accepted in v0
  • < 0.70 — not published in v0
• Confidence is not a calibrated probability. The v0 values predate a formal calibration study. Read the cited evidence andlastVerified date rather than treating a value like0.95 as a 95% truth probability.
• Performance comparisons excluded from v0 because benchmark numbers depend on prompt format, decoding strategy, evaluation harness version, and shot count. Six dimensions of methodology drift make any single “model X scores Y on benchmark Z” claim unreproducible. A future methodology version adds them back with explicit benchmark-version + prompt-format metadata bundled into the envelope.
• SourceScore-issued HMAC-SHA256 integrity metadata. The HMAC tag is not independently verifiable by public users because the shared secret is not published. Public authenticity relies on HTTPS and refetching the canonical record and cited evidence.

Machine-readable methodology + tier reference + endpoint index: /api/v1/methodology.json. Full developer docs (curl + JS + Python examples): /docs/.

Grade scale

Scores 0–100 map to letter grades on an academic-style scale: A+ ≥ 95, A ≥ 85, B ≥ 70, C ≥ 55, D ≥ 40, F < 40. Grade letters are intentionally familiar so the meaning is obvious to a reader who has never visited the site before. Per-grade source rankings: A+, A, B, C, D, F — or see the grading-scale overview.

Current limitations (honest)

• v0.1 publishes 130 hand-scored sources; production scales to 10,000+ via the same rubric.
• Velocity scores are static estimates in v0.1. A future version will refresh Velocity weekly via tier-1 referrer + LLM-citation polling.
• Discipline scores are domain-level. Per-author Discipline (relevant for platforms like Medium) is on the v0.2 roadmap.
• Methodology v0.1 weights are unverified — we will calibrate against actual LLM-citation outcomes after the first 30 days of operator usage and re-tune in v0.2.

No fabricated data

Every numeric score traces to a specific signal a researcher can re-derive (regulatory filings, public ethics codes, academic indexes, structured-data audits, citation-corpus data). Sources whose data we cannot verify are excluded rather than fabricated.

Versioning

The methodology is semver-tracked. Major bumps (vX.0) change scoring weights or add sub-scores; minor bumps (v0.X) refine signals or add sources. Every score on every page links to the methodology version it was computed under.