Skip to content

Evidence retrieval receipts ​

New intelligence and editorial generation uses mandatory Docdex retrieval with dated, source-linked evidence. Connections are modeled possibilities unless the source establishes the relationship.

Authenticated endpoints use the same API key and quota rules as source intelligence:

  • GET /v1/intel/retrievals?artifactId=<artifact-or-digest-task-id> returns related retrieval IDs.
  • GET /v1/intel/retrievals/<retrieval-id> returns search queries, timestamp, index freshness, selected record references, limitations, output hashes and artifact links. Source excerpts are withheld.

A successful empty Docdex search is recorded. An unavailable Docdex service causes retry rather than ungrounded generation. Restricted country-watch/trade-flow rows are excluded from public content, even when an authenticated internal endpoint can access them.

Editorial baselines and scenarios ​

New articles, opinions, investigations and country analyses prepare an entity dossier before writing. Receipt detail includes editorialDossiers, containing dossier/snapshot IDs, prior per-entity snapshot references, source metadata, observation/comparison counts, limitations and three horizon statuses. Unpublished assessment prose and source excerpts are withheld from this endpoint.

The dossier compares dated source-reported conditions, not an assumed complete picture of an entity. It retrieves earlier public source/company records and previous snapshots, excludes AI digests from factual baselines, and refuses future/late-collected evidence when an historical cutoff is requested. Historical excerpts used by a published piece become retained citations.

Horizon definitions are short (0–3 months), medium (3–12 months) and long (1–3 years), measured from dossier asOf. A supported horizon requires assumptions, an alternative, an observable trigger, affected entities/assets, uncertainty and citations. Otherwise it records insufficient_evidence. These are conditional scenarios, not guaranteed forecasts or calibrated probabilities. A baseline review and one bounded repair run before writing. Publication then requires joint evidence review of the dossier and prose. baselineReviewApproved reports pre-writing review status, not publication approval.

Retrieval is bounded: at most 12 entity names and 24 source excerpts, at most 1,200 characters each and 24,000 excerpt characters in total. Measurements must share metric, unit, period duration and scope to support a comparison. A changed retrieval set does not by itself establish an economic change. Sources older than 90 days are flagged as historical; this is a disclosure heuristic, not an expiry rule.

Existing publications remain unchanged and permanently accessible.

Financial baselines resolve the latest available measurement by company, metric and reporting period rather than treating search rank as recency. Structured SEC/FRED metric, unit, period and scope fields are copied by the server; the model selects numbered evidence excerpts. Historical periods may be nonadjacent, and missing current data is an explicit limitation.

For structured SEC/FRED data, current measurements and arithmetic comparison descriptions are server-generated from the source values and literal reporting periods. The AI supplies interpretation, relationships and scenarios. Previous snapshot history is kept separate from the current-condition summary, and a newer company/metric lookup takes precedence over stale search nominations.

The Neural Ledger API