CAPAS compiles supplied evidence into a claim admissibility decision. It does not determine truth. It returns ACCEPT, REWRITE, REJECT, or HOLD with a re-derivable audit trail and no language model in the verdict. First market: statistical claims in regulated submissions.
CAPAS checks whether supplied evidence is internally consistent and re-derivable. It does not certify scientific truth, compliance, authorship, or data authenticity. A consistent fabrication passes: the GIGO ceiling.
python3 benchmarks/home_gate_decisions.pyCorpus: 233 research claims written by AI sessions under human governance in a sister project’s repository. 50 were excluded as contaminated because the coder could see the outcome; 183 were scored. Each claim carries the file and commit where it was asserted. A separate session coded the claims blind. The engine was frozen at a git tag before coding and the verdicts were sealed by hash before the outcomes were opened. No language model in the verdict. Outcomes are the project’s own later corrections, not an external oracle. n is small and the intervals reflect it.
A third party can re-derive these figures from the public files: sealed verdicts, outcome labels and evidence fields in research/ginibre_corpus/public. Each claim text is published as a sha256 hash only, because the paper is unpublished. Replayed with the engine at each frozen tag (ginibre-ciego-v1 to v6), the redacted payloads reproduce 232 of 233 sealed verdicts. Not checkable by a third party: the sealing times and the tag dates, which the author recorded. The full ledger, the limits and what is not validated: CAPAS in depth →
From claim draft to admissibility decision in one deterministic pass.
Every gate run returns exactly one of these. Each carries a machine-readable reason trail.
All schema constraints pass. The declared evidence licenses a statement generated from the declared fields; the original sentence is never carried. No licensing or reproducibility blockers.
Evidence is present but the claim overreaches. Direction, scope, or causal language must be corrected before resubmission.
Critical evidence is missing, contradicted, or irreproducible. The claim cannot be licensed in its current form.
Evidence is incomplete or no external oracle is available; the gate returns HOLD.
Each verdict re-derives to the same audit hash on any machine; no model score is involved.
Not a fact-checker. CAPAS validates evidence licensing and schema compliance, not truth.
Not a plagiarism detector. CAPAS checks admissibility, not citation matching.
Not an LLM wrapper. The decision path makes no language model calls.
Not a replacement for peer review. CAPAS runs before editorial judgment, not instead of it.
A deterministic claim admissibility gate. The same input produces the same output.
A schema enforcement layer. CAPAS Schema v3 defines a valid evidence package.
A provenance audit tool. Every blocker is machine-readable and exportable.
An API for research pipelines, publishing workflows and audit systems.
CAPAS produces an auditable evidence trail for each decision. Every verdict carries an action: ACCEPT licenses reuse · REWRITE returns the corrected claim with an Original→Licensed diff · REJECT names the missing evidence fields · HOLD lists the obligations to resolve before reuse.
Load a sample payload or build an evidence contract. The pilot requires no account.