Function evaluate_retrieval 

pub fn evaluate_retrieval(
    cases: &[EvaluationCase],
    predictions: &[RankedPrediction],
    cutoffs: &[usize],
) -> Result<EvaluationReport>

Scores ranked evidence retrieval without an LLM judge.

The same metrics are usable for LoCoMo dialog evidence, LongMemEval session evidence, and coding-agent provenance. Missing predictions count as misses; unknown and duplicate case identifiers are rejected.

Errors

Rejects empty cutoffs, duplicate identifiers, empty relevance sets, and predictions for unknown cases.