Pre-Registered ML Manifests — an open, content-addressed format for committing a machine-learning evaluation claim to a SHA-256 hash before the experiment runs. Lock the threshold before the data, or it didn't happen.
A verifier with the manifest, the dataset, and the model independently recomputes the digest and emits a deterministic verdict: PASS / FAIL / TAMPERED. Tampering produces a detectable hash mismatch.
- Spec: spec.falsify.dev/v0.1 — CC BY 4.0, v0.2 frozen
- Site: falsify.dev · Registry: registry.falsify.dev
- Citable: Zenodo DOI 10.5281/zenodo.20177839
- Maps to EU AI Act Articles 12/15/18/50, NIST AI RMF, ISO/IEC 42001
- 4 reference implementations (Python, JavaScript, Go, Rust) — byte-equivalent across 20 conformance vectors
- MIT-licensed code · CC BY 4.0 spec · JSON Schema in SchemaStore
- falsify — spec + Python reference + conformance suite (Go/JS/Rust under
impl/) - falsify-js — JavaScript reference (npm:
falsify-js) - mlflow-falsify — MLflow plugin: auto-tag every run with its PRML manifest hash
- prml-verify-action — GitHub Action: gate CI on tampered/regressed eval claims
- falsify-inspect — adapter for Inspect AI eval logs
- falsify-cookbook — patterns, anti-patterns, worked examples (CC0)
- falsify-integrity-index — public scorecard of how published ML eval claims meet falsifiability criteria
Maintained by Cüneyt Öztürk · hello@falsify.dev