Skip to main content
This quickstart traces one hypothetical candidate through every stage of EMEP so you can navigate the specification in order. No code runs; this is a guided tour of the primary research loop and the pages that own each step.

The 8 Stages

1

Discover and import a base model

ModelLoader fetches weights, tokenizer, and configuration from the source. Every artifact is hashed. Read Model Import.
2

Analyze compatibility with a second parent

ModelCompatibilityAnalyzer verifies architecture, tokenizer, tensor shapes, and precision. Emits COMPATIBLE, CONDITIONALLY_COMPATIBLE, or INCOMPATIBLE. Read Model Compatibility.
3

Select a merge strategy and execute

MergeEngine runs the selected strategy (Linear, SLERP, TIES, DARE, DARE+TIES, Task Arithmetic, Franken-Merge). Read Merge Engine.
4

Validate the merged artifact

Numerical checks, dtype checks, integrity hash. Read Merge Validation.
5

Register as a candidate

ModelRegistry assigns an ID and links parent lineage. Read Model Registry.
6

Evaluate on the optimization split only

BenchmarkEngine runs the profile. Hidden Test Set is never touched here. Read Evaluation Framework.
7

Compute fitness and update population

FitnessEngine aggregates metrics. EvolutionEngine ranks, selects, mutates, crosses over. Read Evolution Engine.
8

Promote, quantize, sign, and deploy

Best models are quantized, packaged, signed, and moved through staging to production, including the offline enterprise path. Read Deployment and Offline Deployment.

Data Isolation Guarantee

The EvolutionEngine never sees the Hidden Test Set. This is enforced at the data access layer. Read Benchmark Integrity.

Reading Order

If you set architecture

System, components, runtime, storage, GPU, security, observability.

If you implement merges

Strategies, tensor operations, math, validation.

If you design experiments

Genome, population, selection, fitness, multi-objective search.

If you run production

Runbook, monitoring, alerting, disaster recovery, GPU health.