`pursue-records.csv` — provenance and column notes

478 rows, one per file released by the US Department of War's PURSUE database (Presidential Unsealing and Reporting System for UAP Encounters), collected 2026-09-20 for Task 10 of the timeline plan (.superpowers/sdd/timeline-plan-2026-09-20/task-10-brief.md), and brought up to the seventh tranche on 2026-10-23.

Provenance and licence

Source: war.gov's PURSUE UAP database, published by the US Department of War with ODNI support, following a presidential directive of 19 February 2026. Seven tranches to date at the last collection — first released 2026-05-08, seventh 2026-10-23.
Licence: US government works are public domain. Unlike the NUFORC mirror (see C5 in docs/timeline-feature-2026-09-20.md), nothing here restricts republication.
robots.txt: war.gov/robots.txt was checked before collection and disallows only admin, binary and infrastructure paths — nothing restricts the UAP database pages. No crawl-delay is set. Re-check before any further collection, in case the policy changed.
Collection date: 2026-09-20. 447 rows were observed in the database at that time, matching the count reported by war.gov itself.
Seventh tranche, 2026-10-23: 31 rows taken from the data file war.gov/UFO itself loads (Portals/1/Interactive/2026/UFO/uap-data.csv, which answers a browser but refuses plain curl), 478 rows in all. 28 are release 7. The other 3 (LLE-UAP-D002 to D004, transcripts of the Colorado police videos) carry release 6's date but were not in the database when it was first collected; they are filed under release 6, as the source dates them. All 31 are PDFs. The source's TAG:AAWSAP pairing is expanded to every other AAWSAP-tagged file, as earlier tranches were, so the 44 existing AAWSAP rows also gained the 9 new ids.
What's excluded: the underlying PDFs, videos and audio are not mirrored into this repo — source_url points at the individual record or its download on war.gov, so a reader reaches the actual document there. Mirroring hundreds of files into a static site's git history would be absurd regardless of the licence question.

Columns

Column	Notes
`file_id`	The agency's own identifier, e.g. `DOW-UAP-D079`, `FBI-UAP-PR005`. Not globally unique on its own — see "ID disambiguation" below.
`title`	As given by the source.
`agency`	Contributing agency: Department of War, AARO, FBI, CIA, NASA, ODNI, DoE, State, or a named local law-enforcement body.
`release_number`	Which tranche the file came out in (1–7 at the last collection).
`release_date`	Date war.gov published the release.
`incident_date`	As the source gives it — exact (`7/2/52`), month-only (`OCTOBER, 2023`), or `N/A`. Preserved verbatim; not normalised.
`incident_date_iso`	A normalised ISO date or month derived from `incident_date`, empty where the source gives no usable date. This project does not invent precision the source doesn't have — the same discipline that put the 1990 split in the density ridge (see `scripts/build-timeline.mjs`).
`incident_location`	As given by the source.
`file_type`	`.PDF` / `.VID` / `.AUD` / `IMG`, etc.
`source_url`	Points at the individual record or its download on war.gov.
`description`	The source's own description field, where provided.
`thumbnail_url`	Thumbnail image URL, where the source provides one.
`related_files`	Other `file_id`s the source itself cross-references as related to this record. See "Broken references" below — this column is reproduced as the source wrote it, not repaired.
`virin`	The DoD Visual Information Record Identification Number stamped on the file, where present.
`featured`	Set when the source's own UI flagged the record as featured.
`redacted`	Set when the source marks the record as redacted.

Titles that are filenames

36 rows carry an archive filename as their title (e.g. 18_100754_ General 1946-7_Vol_2, 65_HS1-834228961_62-HQ-83894_Section_001). The CSV keeps them as released. The site shows a readable title instead, written from the record's own description and filename, one per file_id in pursue-title-overrides.csv; scripts/build-records.mjs applies it, and scripts/test-records.mjs fails on any title still containing an underscore. The findings analysis (build-findings.mjs) reads the source titles, because its title-year check is a finding about the government's own wording.

Dates as shown

incident_date is written the US way (7/2/52, June 3-7, 1965, October, 2023). The site shows it day-month-year (2 July 1952, 3–7 June 1965, October 2023), rewritten by displayDate in scripts/build-records.mjs; a two-digit year takes its century from incident_date_iso. Vaguer forms (1958, Late 2025, 1952-1953) are shown as written. The CSV is unchanged.

ID disambiguation (`-R03` / `-R04` suffix)

A small number of file_ids collide across releases — the source reused an id (or a near-id) for what is, on inspection, a distinct file. Where this project could not resolve the collision by using the source's own id as-is, it appended -R0 (the release number) to keep the two rows distinct in this CSV — for example FBI-UAP-D014-R03 and FBI-UAP-D014-R04. This suffix is this project's own local disambiguation convention, not part of the source's identifier scheme. A reader matching ids back against war.gov's own database should strip a trailing -R0 if the bare id doesn't resolve there.

Broken `related_files` references

196 of 2,285 related_files references (8.6%) resolve to no row in this file. This project found three distinct causes, all on the source's side, and did not repair any of them:

Short forms — a reference like PR-019 where the actual id is the fuller DOW-UAP-PR019 (or similar prefix truncation).
Case mismatch — a reference like DoW-... where this file's own id for the same record uses DOW-... (or another capitalisation difference).
Human-readable titles in place of ids — a reference such as Western US Event or ODNI-UAP-D001, USPER Narrative, Senior USIC Official where the source wrote a description or partial title instead of a bare id.

These are not fixed here. related_files is reproduced exactly as war.gov's own database gives it. Rewriting a reference to make it resolve would mean this project inventing a link the government's own referencing style does not actually assert — the same reasoning that keeps incident_date unnormalised and incident_date_iso empty rather than guessed. A reader who needs a broken reference resolved should treat it as a research question, not a data-entry bug.

pursue-records.csv — provenance and column notes