Locked & verifiable.
An evaluation claim anchored to a SHA-256 hash at the recorded time. Anyone can re-derive it from the canonical bytes below.
464df2bee6b479b742306cca356ce8e4904be22cad09beb5cfee29025c1b1d31version: prml/0.1 claim_id: 01a0ce19-5532-7c91-bed6-1ac1d4c0d689 created_at: '2026-09-23T11:49:15Z' metric: accuracy metric_args: split: stratified 70/30 train/test at the manifest seed features: all 20 attributes; categorical one-hot encoded; numeric standardised model_config: scikit-learn LogisticRegression, L2 penalty, C=1.0, max_iter=2000 definition: share of correct predictions (good/bad credit) on the held-out test split comparator: '>=' threshold: 0.7 dataset: id: uci-statlog-german-credit hash: b21f3d81db8071257d5ff1deaeba1fd4303b62712e6fcc9715c7a86202cb5871 uri: https://archive.ics.uci.edu/dataset/144/statlog+german+credit+data seed: 42 producer: id: acceptance-record-example.falsify model: id: sklearn-logistic-regression-l2-C1.0 notes: 'Illustrative acceptance record (criterion C2 of plan AP-2026-001). Falsify OU plays both supplier and client roles; not a client engagement.'raw bytes →
[](https://registry.falsify.dev/464df2bee6b479b742306cca356ce8e4904be22cad09beb5cfee29025c1b1d31)- uses: studio-11-co/prml-verify-action@v2
with:
mode: verdict
expected-hash: 464df2bee6b479b742306cca356ce8e4904be22cad09beb5cfee29025c1b1d31github.com/studio-11-co/prml-verify-action →Verify this hash yourself
Paste your manifest YAML. The canonical hash must match 464df2bee6b4…. This runs the registry's own canonicalization module (canonical.js) in your browser; for verification that does not trust this registry at all, use any of the four reference implementations offline.
Verify the independent timestamp (RFC 3161, offline)
The token countersigns this manifest hash with the timestamp authority's key, so the time claim no longer rests on this registry. Verify with OpenSSL 3 (LibreSSL, the macOS default, cannot check the ESS extension):
curl -sO https://registry.falsify.dev/464df2bee6b479b742306cca356ce8e4904be22cad09beb5cfee29025c1b1d31.tsr
curl -s https://timestamp.sigstore.dev/api/v1/timestamp/certchain -o chain.pem
awk 'split_after==1{n++;split_after=0} /END CERTIFICATE/{split_after=1} {print > ("tsa-" n ".pem")}' n=0 chain.pem
openssl ts -verify -digest 464df2bee6b479b742306cca356ce8e4904be22cad09beb5cfee29025c1b1d31 \
-in 464df2bee6b479b742306cca356ce8e4904be22cad09beb5cfee29025c1b1d31.tsr -CAfile tsa-1.pem -untrusted tsa-0.pem
Expected output: Verification: OK.
What this receipt proves — and what it does not
- Proves: these manifest bytes hashed to this SHA-256, and the registry attests (Ed25519 signature over hash + timestamp, key at /pubkey) that it saw the hash at the recorded time.
- Timestamp: where an RFC 3161 token is present (the "Independent timestamp" row above), the time claim is countersigned by a public timestamp authority and verifiable offline; backdating it would require compromising that authority, not just this registry. Records without a token rest on the registry's clock alone until anchored. Where a "Transparency log" row is present, the hash is additionally mirrored to the public Sigstore Rekor log with a verifiable inclusion proof.
- Does not prove: ordering, or the claimed result. A PRML receipt establishes that these criteria existed no later than the recorded time; it does not establish that the evaluation ran afterwards — that is what a
prml-linkage/0record is for — and it never establishes that the result is good. - The signature covers hash + timestamp only, not the submitted handle. There is no key-rotation story yet; the current key id is served at /pubkey.
- Limitations: the registry itself is a signed key-value store, not an append-only log — per-record non-equivocation comes from the Rekor entry, and records committed before full-manifest storage (2026-07-11) carry their RFC 3161 token only. No long-term storage guarantee: mirror your hashes in your own infrastructure.
PRML v0.2 is a frozen RFC (comment window closed 2026-05-22) — spec.falsify.dev/v0.2-rfc. Editor: spec.falsify.dev/editor.