Sentinel Intelligence

Here is why you can check us

Sentinel is a public map of U.S. political finance built from federal filings. Journalists and staffers need a source that survives a second look, not a scoreboard of totals they cannot check.

The claim we will stand behind
  1. 1.What we show you is what we ingested.
  2. 2.We can prove when we ingested it.
  3. 3.We can prove we have not altered it since.
  4. 4.We cannot prove the filer told the truth.

Anchoring and hash provenance prove a record existed at a time and has not been altered since. They do not prove the underlying federal filing is accurate. That limit is part of the claim, not a footnote.

Coverage dates vary by source. The graph is built from federal filings on an ingestion schedule, not in real time, and some series are months behind their sources. We publish what we ingested and when.

Per-source coverage table →

What a reader can check

Every answer is designed to print two identifiers: a Sentinel trace and the primary-source transaction id where one exists. The second line lets a reader go to the FEC (or other agency) and cut Sentinel out of the loop.

Sentinel trace sentinelintel.org/t/…
FEC transaction 40213…

Trace pages are designed as permanent handles: when one is issued for a claim unit, if the underlying claim is retired, the trace is intended to resolve to the retraction rather than a 404.

Retraction

A headline metric we retired

What it was. We published a capture-classifier performance figure, an AUC-ROC near 0.96 , as if it measured held-out detection of influence.

Why it was circular. The positive label was two input features combined with a logical AND. The model was recovering a definition it had been handed, grading its own homework, not measuring capture on unseen ground truth.

What replaced it. That performance figure is retired. What we publish instead is source-bound: which federal filing a record came from, when it was ingested, and that the stored artifact has not been altered since. No replacement AUC is offered.

What we will not claim

  • -“Blockchain-verified political data” or any formulation that treats a hash anchor as proof of correctness.
  • -Unprovenanced graph totals (entity or edge counts without population, computation date, exclusions, and non-claims). Under standing rule those are removed, not silently corrected.
  • -Named sitting legislators as promotional examples. Tools that answer a user's question are one act; campaign-style name-dropping is another.

Public integrity tools keep going dark

Risk-averse readers need a source that remains checkable when others disappear, not a larger unverifiable map.

Sunlight Foundation, Dissolved
2020
MapLight, Shuttered
2022
ProPublica Represent, Discontinued
2023
OpenSecrets API, Discontinued April 2025
2025

Sentinel is built by one person, from public federal data, under an MIT license. The work is nonpartisan and non-accusatory: connections are mapped; guilt is not assigned.