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.