Evidence performance

Evidence records a derivation for each selected (statement, focus) pair. That work costs more than deciding conformance alone, especially when many nodes pass and only a few failures need explanation.

Choose the smallest result you need

Need	Entry point on a prepared evidence session
Only conformance counts	validate_conformance()
Which selected pairs failed	find_failures()
Derivation for one selected pair	explain(pair)
Passing and failing derivations for every pair	validate()

On measured Brick models, find_failures() followed by explain() for every failure cost 3–34% more than conformance alone. Materializing evidence for every selected pair cost 2.5–5.4 times as much. Those ratios reflect the measured corpus and its share of failing pairs; use them to choose an interface, not to predict a new dataset’s runtime.

What adds cost

The evidence pass retains paths, values, and nested judgments that conformance can discard. Terms and path-support certificates recur in serialized runs. to_compact_json() stores repeated terms and subtrees once; it reduced the measured serialized size by 76% on Brick and 50% on ASHRAE 223P. It does not reduce the work needed to construct the evidence.

Reuse a prepared schema across data snapshots to avoid paying compilation cost for every graph. For large runs, find failing pairs first and explain only those the application will show. See Explain why a node passed or failed for the calls and Evidence reference for their return values.

Measurements and reproduction

The Evidence performance study records the models, timings, memory and serialization measurements, rejected optimization hypotheses, and commands used to reproduce them. Benchmarks measures the separate end-to-end validation pipeline, including fixed setup cost.