Required inputs
Pseudonymous task IDs
Locked predictions
Outcome records
Submission deadline
Join key
Ordered execution
Work the protocol
- 1
Generate task IDs without outcome labels.
- 2
Submit predictions before availability.
- 3
Seal prediction digests.
- 4
Adjudicate outcomes independently.
- 5
Join only by the frozen pseudonymous key.
Outputs
- Blinded task set
- Submission receipts
- Paired scoring table
Done only when
- Every included prediction predates outcome.
- Join cardinality is explicit.
- Excluded rows retain reasons.
Fail closed
Refusal conditions
- Predictor sees outcome or proxy.
- Outcome reviewer sees preferred model label.
- Unmatched tasks are silently dropped.