EMEP Executive System Dependency Map

High-level dependency chain from base models through merge, evolution, evaluation, and deployment, with optional fine-tuning and RAG side loops.

The executive dependency map shows the top-level data and control flow through EMEP. Each stage depends on the successful completion of all upstream stages.

Dependency Chain

Stage Descriptions

Stage	Input	Output	Failure Mode
Base Models	External checkpoints (Hugging Face Hub, enterprise mirror, local path)	DISCOVERED models in ArtifactStore	Import failure, corrupted weights, unsupported format
Compatibility Analysis	Model pairs or sets	COMPATIBLE / CONDITIONALLY_COMPATIBLE / INCOMPATIBLE verdict	Shape mismatch, tokenizer vocabulary mismatch, architecture divergence
ModelRegistry	Validated models	REGISTERED entries with metadata and lineage	Duplicate registration, metadata validation failure
MergeEngine	REGISTERED models + MergeStrategy	Merged weight tensors	Out of memory, numerical instability, strategy mismatch
CandidateGenerator	Merged weights	CANDIDATE model with manifest and signature	Manifest generation failure, signing error
EvaluationEngine	CANDIDATE + benchmark splits	Metric collection per split	Benchmark execution error, timeout, contamination detected
FitnessEngine	Evaluation metrics	Scalar or vector fitness score	Missing required metrics, normalization failure
EvolutionEngine	Population + fitness scores	Next generation or termination	Premature convergence, diversity collapse, GPU exhaustion
Model Selection	Pareto front	Selected PROMOTED model	No non-dominated candidate meets thresholds
QuantizationEngine	PROMOTED model weights	Quantized weights (int8, int4)	Accuracy degradation beyond tolerance, unsupported ops
DeploymentManager	Quantized model + config	RELEASED / DEPLOYED artifact	Packaging failure, signature verification failure, target environment mismatch

Optional Side Loops

• Fine-tuning: LoRA or QLoRA adapters may be applied to base models before merge, producing task vectors. This loop feeds into CandidateGenerator but is not required for the primary pipeline.
• RAG: Retrieval-augmented generation architecture may be attached to deployed models. This is a downstream serving enhancement, not a merge-time dependency.

Feedback Loop

The evolution cycle creates an iterative dependency: EvolutionEngine feeds back to MergeEngine when convergence is not reached. This loop continues until the population converges or a generation limit is exceeded.

Next Steps

See Component Dependency Graph for the internal runtime call graph, or Data Flow Map for the data class flows between stages.