What this actually is, precisely

52,408 trials registered on ClinicalTrials.gov as terminated, withdrawn, or suspended were machine-classified into a 19-code taxonomy describing why each one actually stopped — separating scientific failure from operational reasons (funding, staffing, business strategy, external events) and outright non-failures. These are the dated source corpus and taxonomy claims; importing the snapshot does not repeat its completeness checks.

No human-labeled ground truth exists yet. Every figure above is machine classification, cross-model agreement, or model-adjudicated — never validated accuracy. This site does not round that off.

The source reports cross-model agreement, model self-consistency (91.6% code-level agreement with itself, n=1400), and two checks reported to converge on the same number. These are source-reported checks, not a new validation by ProofPrism. Two independent methods — CT.gov posted-results data and blind Opus adjudication — converge on ~0.2% error for non-contested records.

The dated source reports that records where the classifier's secondary code disagrees with its primary code have a 44% false-positive rate under blind adjudication. Non-contested records run 0.2%. Every headline number on this site excludes contested records for exactly this reason.

Case details and verdict codes are reproduced from imported pilot records. ProofPrism does not independently re-perform or validate the underlying searches during generation.

The source narratives can describe web searches, publications and conclusions. They remain historical source statements, not current verification by this site. A missing pilot observation date stays unknown. A generation timestamp does not establish when a search took place.

The dated source describes 7 audit rounds of its classification pipeline. Several were rounds of fixing real bugs the project introduced and then caught itself — a false-positive regex that flipped real records the wrong way, a re-classification prompt that led its own witness and overstated an error rate by ~25%, a cost estimate off by 3x from unverified hardcoded pricing. 22 corrections were later found wrong and retracted. All of that is tracked in the source repository's commit history, on the theory that a verification project that hides its own errors isn't one worth trusting.

It is an imported qualitative judgment whose authorship this site does not authenticate, not a statistical estimate and not a percentage. Where a numeric probability would imply false precision, this site deliberately does not compute one.