Evidence¶

Evidence records a reusable, source-bound evaluation of a real Computation output. It is not a metric type and it does not replace application control flow: an evaluator reports whether one explicit rule passed, failed, or could not be evaluated.

Declare an evaluation rule¶

Use @evidence for a rule that accepts exactly one value and returns "pass", "fail", or "error":

from oclp import JsonArtifact, computation, evidence


@evidence(name="Validation quality")
def validation_quality(metrics: dict[str, float]) -> str:
    return "pass" if metrics["rmse"] <= 250 else "fail"


@computation(
    name="Evaluate candidate",
    outputs={"metrics": JsonArtifact(name="Validation metrics")},
    requires=(validation_quality,),
)
def evaluate_candidate() -> dict[str, object]:
    return {"metrics": {"rmse": 212.4}}

The SDK publishes one Evidence record for the evaluator's selected output. A returned "fail" or "error", and an evaluator exception, produce a compact Diagnostic. The original Execution and its outputs are still published.

Required Evidence controls the Execution result¶

Every evaluator in requires= runs before the Execution's terminal Event. An Execution can be succeeded only when every required evaluator has Evidence with outcome="pass"; otherwise its terminal status is failed.

The surrounding run continues by default. To stop application control flow after the failing Execution has been completely published, configure the workflow with required_evidence_policy="raise". See Runs for that run-level policy and Computations and Executions for declared output ports and parameters.

When to use Evidence¶

Use Evidence for a durable conclusion about a record: validation quality, schema conformance, a promotion gate, or a safety check. Keep ordinary numeric measurements in an output Artifact; an optional integration such as MLflow can then mirror selected values for comparison without changing OCLP provenance.