◆ Sources · Methodology
Most UAP-related releases live in silos — NARA, AARO, civilian sighting databases, independent FOIA archives. For every document in this archive, we look across those four sources and surface matching records on the document's page. Below: what the four sources are, what we promised them, and how the matcher scores a pair.
National Archives Record Group 615 — canonical UAP collection
US-government records, public domain. NARA publishes a structured bulk-download manifest (S3 + JSON) that we mirror metadata-only. We never republish full PDFs from NARA — every match links to catalog.archives.gov where the canonical PDF lives.
Today · 35 record-series ingested. Series-level only for now; per-item NAID ingest is gated on getting a NARA Catalog API v2 key.
All-domain Anomaly Resolution Office — DoD case resolutions
US-government case-resolution PDFs from the All-domain Anomaly Resolution Office. Public domain, but aaro.mil hard-blocks programmatic fetch (Akamai bot wall on every UA/header combo). Our index is hand-curated from public reporting + congressional-hearing coverage. Matches link to the source PDF on aaro.mil.
Today · 7 case resolutions indexed. PDF bodies not OCR'd at ingest time — title + date + location drive the matcher.
National UFO Reporting Center — civilian witness reports
Civilian-witness sighting database with an explicit no-scraping ToS clause. We do not maintain a local cache and we never excerpt witness narratives. Matching against NUFORC is gated behind an explicit operator opt-in (NUFORC_FEDERATION_ENABLED=1) so a fresh checkout doesn't quietly federate. When enabled, the matcher hits NUFORC's own public search URLs at sidecar-build time and stores only the resulting outbound link. If NUFORC grants a written research-archive exception, that's the path to broaden the integration.
Today · 0 matches surfaced. Live-federation disabled by default.
The Black Vault — independent FOIA archive
Independent FOIA archive curated by John Greenewald. Underlying documents are public domain (FOIA releases) but the curation, OCR, and editorial annotation are Greenewald's work. We render outbound link + title + date only — no body excerpt — until an attribution exchange is on file. robots.txt explicitly allows our crawler, so the index pull is sanctioned.
Today · 15 prominent posts indexed by hand. Future upgrade: WP REST or sitemap-driven crawl.
Every (our_document, foreign_record) pair gets a score in [0, 1] from a weighted sum of four components. Weights are renormalized while the embedding-similarity component (worth 0.20 in the long-term design) is still offline.
Pairs scoring ≥ 0.50 surface as definite matches by default. Pairs from 0.40 to 0.50 are possible matches, tucked behind a disclosure on each doc page. Below 0.40 the matcher discards the pair before it reaches the sidecar.
The matcher is heuristic — it errs on the side of recall in this MVP. If a definite match on your document is wrong, or a source you maintain wants its material handled differently, we'd like to know.
The full design and per-phase changelog lives in
docs/superpowers/plans/2026-05-19-sprint-d.md
in the project repo.