Verification & Receipts · Simulation
Every result is checked
Pick a submission and see each verification step, the numeric tolerances and the reason code when something fails.
Choose a submission
- Order
- demo-ord-0001
- Task
- demo-task-0001
- Lease
- demo-lease-0001 · 120s
- Submitted
- 21s after lease start
- Records
- rec-001, rec-002, rec-003, rec-004
Verification report
- 1Device signaturepassedMatches registered key
- 2Submitted before lease expirypassedOn time
- 3Model profilepasseddemo-profile/minilm-l6-v2/cpu-fp32/r0
- 4Record IDs, count and 384 dimensionspassedMatches
- 5Finite valuespassedNo NaN or infinity
- 6Cosine ≥ 0.9999, max error ≤ 0.001, norm 0.999–1.001passedWorst cosine 1.000000, max error 2.00e-6
The first profile uses full-reference checking: a trusted verifier recomputes every record. Reference vectors here are authored fixtures, not real model output. A rejected attempt earns nothing and the customer is not charged; the task is retried elsewhere.
What a receipt cannot prove: hashes, device attestation and on-chain settlement do not by themselves prove that the AI computation was correct. Acceptance rests on the trusted verifier's recomputation.
Review options
Proposed: a contributor can ask for review of a rejection within 7 days, and a customer can dispute delivered results within 48 hours. The operator decides in the first release. This demo does not open a real support ticket.