Sources & Pipeline
Data sources
The pipeline polls a fixed, published set of sources several times a day. Primary and structured sources anchor every entry; the press layer supplies cross-spectrum framing and first notice.
Primary / structured (polled every run):
- Federal Register API: the authoritative record of executive orders, proclamations, and presidential documents: what the administration did, structured and dated.
- whitehouse.gov Presidential Actions RSS: fastest first notice of new executive orders and proclamations, cross-checked against the Federal Register.
- CourtListener API v4: litigation status: filings, rulings, and compliance-or-defiance evidence, the backbone of the courts indicator.
- DOJ press releases RSS: enforcement actions and DOJ's own framing of legal posture.
Press / framing layer (polled every run, for cross-spectrum coverage):
- NPR Politics RSS: institutional-center framing.
- The Hill (Administration) RSS: dedicated administration beat, high-frequency daily coverage.
- Fox News Politics RSS: right-of-center framing, required for cross-spectrum balance.
- Just Security RSS: legal/national-security analysis with strong institutional and exculpatory nuance.
Pattern/discovery layer (polled 1–2x/day):
- GDELT DOC 2.0 API: cross-domain volume and pattern signal; also a safety net that surfaces wire-service-originated stories no longer available via direct RSS.
Sources considered and excluded, with reasons, are documented in the pipeline's internal source inventory: AP and Reuters RSS (both discontinued), Truth Social's API and ICE.gov/DHS.gov feeds (bot- or WAF-blocked to automated access), and Lawfare's RSS (blocks non-browser requests).
Discovery panel (lead generation)
The pipeline also polls a fixed panel of six opinion, analysis, and synthesis publications, once per run alongside every other feed. Their function is narrow and disclosed here in full: they surface candidate stories the primary/structured and press feeds may rank low or miss entirely. They do not, and structurally cannot, source a published entry.
The panel, lean stated plainly:
- Heather Cox Richardson ("Letters from an American"): left-of-center.
- Lucid (Ruth Ben-Ghiat): left-of-center.
- Democracy Docket: left-of-center.
- The Dispatch: right-of-center.
- National Review: right-of-center.
- Washington Examiner: right-of-center.
Three voices lean left of center, three lean right of center. The panel is balanced by construction, not by outcome: it is reviewed for composition, not for how often either side's items happen to seed a published story in a given period.
The lead-only rule. Every item ingested from these six feeds is tagged lead_only in the pipeline's store at the moment it is ingested. A lead-only item may be clustered into a candidate story at triage. That is its entire purpose, but it is never treated as a source. It cannot appear in a published entry's sources list, and no factual claim in a published entry may be attributed to one. The pipeline's synthesis and compliance-gate stages (pipeline/prompts/synthesize.md, pipeline/prompts/gate.md) enforce this as a hard rule: a draft that cites a lead-only item's url, or attributes a claim to one, fails the compliance gate outright and does not publish. Where a lead-only item points to a real development, the underlying facts still have to be established independently from the primary/structured sources, the press layer, or the research pass, before an entry can be written.
Why a discovery-only panel is compatible with a neutral sourcing standard. What gets noticed shapes what gets covered, even when what gets published is sourced and framed neutrally. A discovery layer drawn from one side of the spectrum would bias coverage before a single word of an entry is drafted, regardless of how rigorously that entry was later sourced. The panel is balanced left and right in equal number. That keeps the pipeline's upstream influence even-handed, without importing either side's framing, sourcing standards, or editorial judgment into what gets published.
Provenance is logged. Each pipeline run records, per story, whether any of its clustered items came from the discovery panel and which feed(s) they came from, alongside the panel's per-feed item counts for that run. This provenance record is part of the same run log referenced below.
Backfill / pattern-memory corpus
Before any live entries existed, the store was seeded with non-republished pattern-memory context (tagged backfill): existing democracy-scoring waves (V-Dem, Freedom House, EIU, Bright Line Watch, Levitsky/Ziblatt/Way) and an independently hand-researched, individually-sourced corpus of dated events since January 2025. It also includes the public /feed.xml sample (50 most-recent dated, sourced, domain-tagged actions) from the Trump Action Tracker (Christina Pagel, UCL), credited here as required by its content license, CC BY-SA 4.0; the site's full ~3,500-action archive was not bulk-ingested, since its Terms of Use prohibit automated scraping of the site and its only bulk-CSV export is a human-only, in-browser "Download CSV" action.
Source-selection rules
Sources are chosen to be cross-spectrum and primary-documents-first: a claim is anchored to a government document or court record wherever one exists, and press coverage is drawn from outlets spanning the political spectrum rather than any single framing. Two independent sources are required for contested facts. A primary document is linked inline whenever one exists.
Pipeline run log
No pipeline runs have executed yet. When the first scheduled run completes, this section will list each run's timestamp, how many candidate stories were considered, how many entries published, and the outcome, including "quiet run" outcomes, which are treated as data, not failure.
Corrections policy
Factual errors in a published entry are corrected in place. Corrections are dated and left visible: a strikethrough on the corrected text plus an amendment note explaining what changed and why, never a silent edit. Changes to a specific indicator's standing reading are logged as a dated, append-only changelog entry with a stated rationale on that indicator's own page. Changes to the Indicator Framework itself (definitions, zones, or scoring principles) are versioned and dated on the Methodology page.
(This is the same policy published on the Methodology page; if the two ever drift, the Methodology page's wording governs and this page will be updated to match.)