Provenance & Source Index

Source Collections

Where the knowledge comes from

The current corpus is organised into eight public collections. Counts are documents learned, not files we store. Each collection carries a typical trust tier; individual items can differ and are tiered on their own merits.

T2 · Primary

Witnesses & Whistleblowers

1,298

documents · first-hand testimony

T2 · Primary

Major Historical Case Files

1,075

documents · legacy investigations

T3 · Research

Science & Technology

952

documents · analysis & sensor physics

T1 · Official

Government Documents

740

documents · official releases

T3 · Research

Black Budget

263

documents · funding & programs

T3 · Research

Crash Retrievals

229

documents · recovery accounts

T2 · Primary

The Disclosure Project

222

documents · witness testimony

T2 · Primary

Videos & Photos

120

items · analysed, not embedded

Incoming

T1 · Official

US National Archives

NARA Record Group 615

Queued for staged ingestion · links to catalog.archives.gov

Tiers shown are each collection's typical source reliability, kept separate from an item's individual quality score. A document's own tier can be higher or lower than its collection default.

Anatomy of a Citation

How a piece of evidence appears in an answer

When the AI cites something, it shows you this — an Evidence Knowledge Card. It states what the source is, how reliable it is, how much of the claim is established fact, how confident the model is, and where the original lives.

2004 USS Nimitz “FLIR1” encounter — DoD-released forward-looking infrared video

T1 · Official

Official U.S. Department of Defense video, publicly released and authenticated by the Pentagon. Depicts a sensor track recorded by a Navy F/A-18 during the 2004 Nimitz Carrier Strike Group operations off Southern California. The AI draws on the released footage and its accompanying official statements.

Source type · Government video release T1 · Official Claim level · Known

AI confidence High · 0.86 — release & authenticity established; identity of the object remains unexplained.

Knowledge version corpus 2026.07 · model v3

View original at official archive (DVIDS) →

If no public original exists, the card says so — we never fabricate a source. The link always points out to the official public archive; UAP Intel never serves the file itself.

The claim level marks how much is established: Known (documented fact) · Claimed (asserted by a source) · Inferred (reasoned, not stated) · Unknown (unresolved).

Provenance Metadata

What we record per item

For every piece of evidence the AI learns from, we store a provenance record. This is what lets any answer be traced back to exactly where it came from — without keeping a copy of the source material.

source_collection — which of the collections above the item belongs to
source_url — the canonical link to the original on its official public archive
archive_series — the archive's own series or record-group identifier (e.g. NARA RG 615)
download_date — when the source was last read from its archive
sha256 — content hash captured at read time, so the original can be re-verified against the archive
tier — the T1–T4 trust tier assigned to the item
evidence_type — document, testimony, image, video, dataset, etc.
quality_score — the item's individual quality rating, independent of its tier
processing_version — the pipeline version that ingested and analysed the item
included_in_model_version — which knowledge version(s) learned from this item

UAP Intel does not host the archive. We preserve provenance and link to the official public source. The documents and videos already live on public archives such as catalog.archives.gov and DVIDS — we understand them and point you to them, we do not copy, mirror, or re-serve them.