評価

Layer 2: Compose

Evaluatorとグラフネイティブ契約

graph-native経路は GenomeBatch → EvaluationAdapter → EvaluationPayload → Evaluator → ObservationBatch です。

saealib.Evaluator	Base class for batch evaluators.
saealib.SerialEvaluator	Default evaluator: evaluates each candidate sequentially.
saealib.JoblibEvaluator	Parallel evaluator backed by joblib.
saealib.ThreadPoolEvaluator	:py:class:`~saealib.execution.evaluator.AsyncEvaluator`の別名です。
saealib.AsyncEvaluator	Thread-backed asynchronous evaluator.
saealib.EvaluationAdapter	Transform genomes into the payload consumed by a problem evaluation.
saealib.EvaluationRequest	Input payload whose snapshot semantics depend on how it was built.
saealib.EvaluationHandle	Runtime reference to submitted evaluation work.
saealib.EvaluationQuery	Request context supplied to an EvaluationAdapter.
saealib.EvaluationUpdate	One delivered evaluation update.
saealib.PendingEvaluation	Serializable record for submitted work.

評価ポリシーと計画

saealib.EvaluationPlanner	Public contract for planners that may produce multiple requests.
saealib.EvaluationPlan	Composable evaluation work returned by a planner.
saealib.EvaluateAll	Evaluate every candidate.
saealib.TopKEvaluation	Evaluate the stable top-k acquisition scores.
saealib.RatioEvaluation	Evaluate the stable prefix selected by a ratio.
saealib.RepeatedEvaluation	Request candidates with an explicit replicate number.
saealib.FidelityEvaluation	Attach an explicit fidelity level to an evaluation request.
saealib.FidelityPromotion	Represent an explicit promotion from one fidelity level to another.
saealib.ReplicateSummary	Aggregated observations for one candidate batch.
saealib.aggregate_replicates	Aggregate repeated objective observations by stable candidate ID.

Stage互換型

Stage互換経路では Evaluator.evaluate_batch(...) -> EvaluationResult を使います。EvaluationResult の f、g、cv のshape・行数整合性、および candidate_ids を指定した場合の結果行数との一致と一意性が、この互換境界の制約です。Evaluator.submit で candidate_ids を返す場合は、requestのcandidate IDsと完全一致する必要があります。

saealib.EvaluationResult

Validated numeric result for a batch of candidates.