How we grade conclusions

Visible edition label: 2026-11

Sourcebook entries distinguish what documents report from conclusions we draw. A page has one of two status lines:

There is no pending status. A page with our own conclusion must have a completed assessment before publication. Reporting an author’s attributed judgment does not make that judgment our conclusion.

Probability and confidence

We use the National Intelligence Estimate (NIE) Words of Estimative Probability (WEP) vocabulary. These are approximate probability bands, not measured frequencies or promises of certainty. The bands follow the Office of the Director of National Intelligence (ODNI), Intelligence Community Directive (ICD) 203, Analytic Standards, signed January 2, 2015, as amended. The labels below are the exact labels used in our assessment schema; INSUFFICIENT BASIS is our unexamined-claim category and has no probability band.

Label Approximate probability
ALMOST CERTAINLY 95–99%
VERY LIKELY 80–95%
LIKELY 55–80%
EVEN CHANCE 45–55%
UNLIKELY 20–45%
VERY UNLIKELY 5–20%
REMOTE 1–5%
INSUFFICIENT BASIS N/A (not applicable)

Confidence describes the quality of the evidence and reasoning separately from probability. High means strong sourcing and reasoning, not certainty. Medium means credible sourcing with gaps or incomplete corroboration. Low means fragmented evidence, weak corroboration, or uncertain reliability. A low-probability conclusion can have high confidence when strong evidence weighs against it.

Evidence base

Source tier Meaning
primary-multiple At least two independent primary records, such as sworn testimony, court records, official reports, declassified files, or on-the-record admissions
primary-single One primary source
secondary Scholarship, books, or institutional analysis drawing on primary material
journalistic News reporting or investigative journalism, including unnamed sources
speculative Inference, pattern matching, or an unsourced assertion

Corroboration is independent when sources have separate origins; partial when some support is independent but origins may overlap; single-source when no corroboration is available; and contradicted when credible evidence opposes the claim. Repeating one account across several publications is not independent corroboration.

Our evidence rules require primary-single or stronger sourcing and at least partial corroboration for LIKELY. VERY LIKELY requires primary-multiple and at least partial corroboration. ALMOST CERTAINLY requires primary-multiple and independent corroboration. Secondary sources alone cannot support LIKELY or higher. Evidence floors do not automatically award a grade: the reasoning still has to justify it. A strong label cannot conceal a weak evidence base.

Each block states one central conclusion, its evidence basis, and the assessment date. At LIKELY or higher it also names a specific observation that would revise the conclusion, the strongest counterargument, and our response. “More evidence” is not a checkable revision condition.

When a grade changes

A changed grade requires an Assessment Change Record explaining the old and new judgments, the evidence or reasoning that changed them, and what would change them again. A snapshot preserves the assessment state. Edition release notes list every changed grade and link its change record. Changes to source tier, corroboration, confidence, or conclusion also require a change record and snapshot. An issued edition remains fixed; later editions and corrections preserve the history so readers can see what we knew and why our judgment changed.