A truth layer, not a dashboard.

The problem.

AI agents are graduating from drafting text to making decisions — approving a counterparty, validating a route, classifying goods. The moment an answer has regulatory or financial consequence, “the model said so” is not a provenance story. Retrieval over documents gives agents plausible text; decisions need auditable facts.

The reference truth those decisions rest on — what’s a real port, which jurisdiction is high-risk, who is on which list, what category of goods is moving — is scattered across dozens of sources, each with its own format, freshness, and license. Scientia Basis unifies them into one corpus an agent can query, trust, and cite.

The bet.

The interface layer is being commoditized by assistants; the data layer isn’t.

That’s why there is no dashboard: the UI for a non-technical user is any assistant with the MCP server attached. What can’t be commoditized is a corpus where every value carries its own audit trail.

Everything is a statement.

Most databases store values. Scientia Basis stores claims about values — every fact carries which dataset asserted it, the exact source URL, which ingest run produced it, and when it was last confirmed. No fact is ever deleted: facts that vanish from a source go stale, visibly. Conflicting sources coexist as two statements, never a silent overwrite.

country:US · CountryStatement anatomy

Property

postal_code_regex

Value

^\d{5}(-\d{4})?$

Dataset

sci-bas_countries

Source

sci-bas://country_master

Run

3 · sci-bas_countries

Seen

first 2026-08-08 · last 2026-08-08

Four capabilities, one schema decision.

The pipeline.

Fetch, normalize, record, validate, release — then everything reads from the same query service.

A corpus you can pin.

Updating & audit.

No fact is ever deleted — every value carries when it appeared and when it was last confirmed, and the model extends to full point-in-time queries without a rebuild.

This is the operator view. run is each dataset’s own ingest count and is what a citation carries; run_id is a corpus-wide counter, returned only to a key that already reaches every dataset, because its gaps would otherwise tell a narrower key how many ingests it is not being shown.

$ scientia runs --limit 3 --pretty
{
  "corpus_release": "2026.08.7",
  "count": 3,
  "results": [
    { "run": 12, "run_id": 37, "dataset": "sci-bas_factbook",
      "started_at": "2026-08-09T02:17:28Z",
      "finished_at": "2026-08-09T02:17:47Z",
      "source_version": "sha256:52e1c8aee8ab4d78",
      "snapshot_digest": "52e1c8aee8ab4d78f6fc…",
      "status": "ok", "entities": 248, "statements": 62478,
      "release": "2026.08.7" },
    { "run": 9, "run_id": 36, "dataset": "cldr_names", … },
    { "run": 9, "run_id": 35, "dataset": "iso_country_names", … }
  ]
}

Historical operator output from August 2026. The legacy Factbook dataset shown here is held from external serving; corrected releases use the independently sourced cia_factbook dataset. Its new acquisition does not rewrite this historical run.

Honest screening.

Two historical Yente responses from release 2026.08.6 illustrate transliteration-aware matching. Their different-script queries resolve to the same designation, csl:SDN:8279, with engine-specific scores and explanations. These are recorded examples, not a claim about current matching availability.

Initial launch is pinned to snapshot 2026.10.2 and currently uses the explicitly degraded lexical fallback because the stronger index is unaligned. It does not provide Yente's transliteration capability. Check the current screening limits and reject degraded results when that capability is required.

Screening returns candidates, not verdicts. A score ranks likelihood, the threshold is yours to set, and the record behind a hit is a citable id you can fetch, pin to a release and diff. Nothing here decides whether to file a report — that judgement stays with the person accountable for it, and the point of the explanations is that they can disagree on specifics.

{
  "corpus_release": "2026.08.7",
  "match_count": 1,
  "matches": [{
    "id": "csl:SDN:8279",
    "caption": "BOUT, Viktor Anatolijevitch",
    "dataset": "us_csl",
    "score": 0.864,
    "explanations": { "name_match": 0.864 }
  }]
}

{
  "corpus_release": "2026.08.7",
  "match_count": 1,
  "matches": [{
    "id": "csl:SDN:8279",
    "caption": "BOUT, Viktor Anatolijevitch",
    "dataset": "us_csl",
    "score": 0.909,
    "explanations": { "name_match": 0.909 }
  }]
}

Three doors, one corpus.

All three surfaces are thin doors over one query service — an agent, a pipeline, and a terminal get identical answers with identical provenance.

$ claude mcp add --transport http scientia-basis https://scientia-api-go.fly.dev/mcp \
    --header "Authorization: Bearer $SCIENTIA_API_KEY"
# tools: search_entities · get_entity · screen_name ·
#        get_country_profile · get_country_factbook · list_datasets

$ curl -H "Authorization: Bearer $SCIENTIA_API_KEY" \
    'https://scientia-api-go.fly.dev/v1/countries/VN'
$ curl -H "Authorization: Bearer $SCIENTIA_API_KEY" \
    'https://scientia-api-go.fly.dev/v1/screen?name=Aerocom+SRL&threshold=0.7'

$ export SCIENTIA_API_URL=https://scientia-api-go.fly.dev
$ scientia search "vladivostok" --schema Location
$ scientia country VN            # reference + risk.fatf_list · risk.cpi_score · risk.wgi_*
$ scientia datasets              # license, redistribution term and freshness

Start free. Pay for volume, not for access.

Dataset access and screening behavior follow the entitlement policy configured for the deployment. Plans differ in how much you can spend, and nothing else.

Where this goes.