Signed model evaluation manifest

Hashing or bundling an ONNX artifact proves supply-chain identity, not fitness for identity decisions. Model Foundry accepts production promotion only after a current Ed25519-signed proofline.model-evaluation.v1 manifest is attached to the exact artifact SHA-256.

Required evidence

• intended use and explicit limitations;
• countries, processing regions, document types, devices and attack classes;
• evaluation dataset SHA-256, provenance and licence reference;
• no training/evaluation overlap;
• at least 1,000 genuine and 100 negative samples (300 for PAD, injection, deepfake and document-authenticity capabilities);
• at least three declared slices with 50 samples each;
• capability-specific performance metrics;
• start, completion and expiry dates (maximum validity 400 days);
• two distinct evaluation approvals; and
• a detached Ed25519 signature over canonical JSON payload.

Required metrics

Capability	Metrics
Document authenticity	falseAcceptRate, falseRejectRate
Document extraction	characterErrorRate, fieldAccuracy
Face detection	falsePositiveRate, falseNegativeRate
Face match	falseMatchRate, falseNonMatchRate
Liveness/PAD	apcer, bpcer, acer
Injection defense	precision, recall, falseAcceptRate
Deepfake detection	precision, recall, falseAcceptRate

Rates must be between zero and one. Latency metrics must be positive and no greater than 60 seconds. Production deployment scope cannot exceed the signed evaluation scope, and the person promoting a package must differ from the submitter.

Upload

curl -X POST \
  'https://control.assayra.example/v1/platform/model-packages/mdl_123/production-evaluations' \
  -H 'Authorization: Bearer CONTROL_TOWER_TOKEN' \
  -H 'Content-Type: application/json' \
  -d @signed-evaluation.json

Expired or signature-invalid evidence fails closed. New traffic is not assigned until a replacement manifest is approved; historical inference receipts retain the artifact and evaluation lineage used at decision time.