EMEP Data Flow and Class Flow Map

Data-flow diagram showing Weights, Tokenizer, Config, Task Vector, Metric, Manifest, and Signature flows, with explicit benchmark split boundaries.

EMEP processes seven primary data classes through its pipeline. This page maps each class to its origin, transformations, and destinations, with explicit boundaries between benchmark splits.

Data Classes

Class	Origin	Transformations	Destination
Weights	ModelLoader loads from ArtifactStore	MergeEngine combines via TensorEngine; QuantizationEngine reduces precision	CandidateGenerator, InferenceBackend, DeploymentManager
Tokenizer	ModelLoader loads from ArtifactStore	ModelCompatibilityAnalyzer validates vocabulary alignment	EvaluationEngine (for benchmark tokenization), InferenceBackend
Config	ModelLoader loads from ArtifactStore	MergeEngine produces merged config; QuantizationEngine adds quantization config	CandidateGenerator, DeploymentManager
Task Vector	TensorEngine computes (fine-tuned minus pre-trained)	MergeEngine scales, trims, drops, or interpolates	MergeEngine output weights
Metric	EvaluationEngine computes from InferenceBackend output	FitnessEngine normalizes and aggregates	ExperimentTracker logs, EvolutionEngine uses as fitness
Manifest	CandidateGenerator assembles from merge metadata	ExperimentTracker appends experiment ID and timestamps	ArtifactStore, ModelRegistry
Signature	ArtifactStore computes ed25519 hash	DeploymentManager verifies on release	ArtifactStore, DeploymentManager

Data Flow Diagram

Benchmark Split Boundaries

The diagram shows three distinct evaluation data flows:

• Optimization Set (orange): Metrics feed directly into FitnessEngine and EvolutionEngine. This split is used for hyperparameter tuning and merge strategy selection.
• Validation Set (orange): Metrics are logged to ExperimentTracker for monitoring and early stopping. Not used for fitness computation.
• Hidden Test Set (red): Metrics are computed once per experiment, after evolution converges, and are never exposed to EvolutionEngine or FitnessEngine. This enforces CONST-05 and ADR-006/007.

Flow Constraints

1. Weights flow is unidirectional: Weights move from ArtifactStore → ModelLoader → TensorEngine/MergeEngine → CandidateGenerator → QuantizationEngine → InferenceBackend. There is no reverse flow.
2. Task Vectors are ephemeral: Task Vectors exist only in memory during merge operations. They are not persisted to ArtifactStore.
3. Metrics are append-only: Once computed, metrics are written to ExperimentTracker and cannot be modified. Invalid metrics are marked INVALID, not overwritten.
4. Signatures are computed post-merge: The ed25519 signature covers the merged weights, config, and manifest. It is computed after CandidateGenerator produces the manifest and before DeploymentManager releases the artifact.
5. Hidden Test Set metrics are isolated: The red boundary indicates that Hidden Test Set metrics never enter the EvolutionEngine loop. Access is logged and audited.

Next Steps

See Experiment Lifecycle for the state machine governing these data flows, or System Dependency Map for the executive pipeline view.