How we score, verify, and trust

1. Officialness score

Every candidate API or dataset is scored on a 0–100 "officialness" scale before it can enter the public catalog. The score is the weighted sum of five proof signals:

Entries scoring ≥ 80 with a known institution are auto-approved. 40–79 are quarantined for human review. Below 40, or under a blocklisted domain, the candidate is rejected.

2. Continuous verification

Every active entry is re-checked on a rolling 6-hour cadence — APIs by hitting their declared base or health URL with the appropriate method, datasets by re-checking each declared distribution. Three consecutive failures move an entry from VERIFIED to FAILING; a clean check restores VERIFIED. We treat 401/403 as "up, requires auth" rather than "down" — many official APIs gate behind API keys but stay alive.

Verification evidence is recorded in the checks table: timestamp, HTTP status, latency, a short error snippet on failure — and, on every passing check, the exact request that proved the entry alive. That stored "last known working request" is what the entry page's Run verified request action replays. Uptime history is exposed at /api/v1/entries/{slug}/checks.

3. Access grading (the Access Contract)

Every entry is graded by what you can actually do with it right now. direct: the primary action yields the data itself — a downloadable distribution or a stored, verified request that returns data. documented: machine-readable but indirect access — an OpenAPI spec, a bare base URL, an auth-gated endpoint. reference_only: all we can offer is a landing page. The grade is recomputed whenever new evidence lands (a profile, a passing check) and is filterable across the catalog (?access_level=direct, ?format=CSV). Long-standing reference-only entries are pulled back into human review rather than padding the catalog.

The grade feeds the Trust score shown on each entry — a five-axis, MQA-style composite (officialness, availability, metadata completeness, license clarity, agent readiness) modeled on the data.europa.eu Metadata Quality Assessment ranges.

4. Field-level provenance

Every visible value (name, description, license, base URL, …) carries a provenance row stating which extractor produced it, from which source URL, with what confidence, and a 1 KB evidence snippet. The cascade runs in priority order — OpenAPI specs > repo metadata > HTML docs > heuristics — and we record the winning extractor, not just the value. See any entry's Provenance tab.

5. License normalization

Open-data licenses are a thicket — CC variants, OGL-UK, Licence Ouverte, IODL, DL-DE, ODbL, EUPL, national variants. We normalize raw license strings into a canonical id by pattern-matching against a reference table of 19 licenses, preserving the original string for audit. Each id carries boolean flags for is open, is redistributable, and requires attribution. Unrecognised strings stay raw, never silently mis-classified.

6. Standards we speak

• DCAT-AP 3.0.1 + the W3C DCAT 3 vocabulary: the catalog is served as Turtle (/api/v1/catalog/dcat.ttl) and JSON-LD (/api/v1/catalog/dcat.jsonld), harvestable by data.europa.eu and national portals.
• schema.org/Dataset: every entry page emits a JSON-LD block readable by Google Dataset Search.
• FAIR Signposting: each entry page carries a rel="describedby" link to its machine-readable record.
• OpenAPI 3.1: we publish our own API surface at /openapi.json.
• llms.txt: agent-discoverability for AI clients that prefer the convention over web search.
• MCP: the register is a Model Context Protocol server at /mcp — search, per-entry access contracts, and join-key traversal as tools any MCP client can call.
• Per-entry DCAT: each entry serves its own record (distributions included) at /api/v1/entries/{slug}/dcat.jsonld, and the entry graph is exported at /api/v1/graph/edges.

7. Independence

We are not a national portal. We do not represent any institution we list. Every signal you see is reproducible from primary sources — institutions, portals, distributions — and the methodology is the same regardless of who publishes. If you spot a wrong call, the submission flow takes corrections.