Predictions interface contract
The shared file format used to validate and evaluate predictions from different frameworks.
Glossary
Terms used in the evaluation workflow and reports.
Core terms
Wrapped predictions
A file containing predictions, their schema version, and metadata about how they were produced.
Protocol identity
The recorded dataset, split, metric settings, thresholds, and other evaluation choices that determine whether reports are comparable.
Export settings
Producer-side preprocessing and output settings recorded in meta.export_settings so prediction artifacts can be interpreted and compared.
Stable lane
The supported core features for validating and evaluating predictions.
Bridge lane
An adapter or export workflow that converts external outputs to the shared prediction format.
Benchmark lane
Checks whether runtimes or backends produce equivalent results, after Stable evaluation works.
Research lane
An optional experiment using an evaluated baseline. Its results do not establish production readiness.
Skipped lane
A path that did not execute because a capability, artifact, runtime, or prerequisite was unavailable; it is not a passing result.
Comparable report
A report whose task, data, protocol, predictions provenance, and metric semantics align with the report it is compared against.