About this project & how the data is made

What this is

A private research tool that makes the complete public record of U.S. security-clearance decisions readable: 29,877 decisions — 26,269 hearings and 3,608 Appeal Board rulings — searchable in full, summarized case by case, and counted honestly, with data through March 2026. It was built by a subject-matter specialist in clearance adjudication, for the three groups who need this record: people facing a Statement of Reasons, the attorneys who represent them, and researchers studying how adjudication actually works. It is not affiliated with DOHA, DCSA, DoD, or the U.S. Government.

Where the data comes from

Every decision is a public record published by the Defense Office of Hearings and Appeals. We collect the published files, extract their text with layout-aware tooling, and preserve the original PDF on every case page — so anything we say can be checked against the source in one click.

How summaries and statistics are made — and checked

Machine reading does the heavy lifting; verification is the product. The discipline, in plain terms:

• Grounding checks. Extracted allegation text is verified against the decision’s own words by automated checks. Text that can’t be grounded is replaced with an honest “not stated in the decision” — never invented — and anything below our confidence bar is visibly flagged on the page.
• Sampling audits. Summaries and outcomes are audited line-by-line against source decisions in repeated sampled batches. When an audit finds a failure mode, the affected data is regenerated corpus-wide, not patched over.
• Verified case types. Whether a decision is a hearing or an appeal is read from the decision text itself (the source labels were unreliable), and hearing and appeal outcomes are never mixed in one statistic. In our audit of the appeal corpus, the Board’s stated outcome matched our independent read in 100% of sampled cases.
• Honest denominators. Every rate says how many cases it is based on; small samples are flagged as unreliable rather than displayed as fact; thin years are shown as volume, not noisy percentages.
• AI never rewrites the record. Verbatim legal text is displayed from the source or not at all. The AI-assisted case summaries are built only from separately verified facts and are gated by automated checks before display.

Known limits

The public archive itself is thin for hearing decisions in roughly 2010–2017; published decisions are redacted by the government; and no extraction process is perfect — which is why the original PDF travels with every case, and why we correct verified errors (see Disclaimers).

The rules we operate by

• The public record stays free to read. Always.
• Everything is descriptive — never a prediction, score, or advice.
• Uploads are private, encrypted, deletable, and auto-expiring (see Privacy & Security).
• When we’re not sure, the page says so.