Mutation is a declared transformation
A mutation operator maps one or more parent packages to a descendant. Operators should be allowlisted, versioned, deterministic where possible, and constrained by explicit parameter ranges.
Model-level operators
| Operator | Best use | Main controls |
|---|---|---|
| Fine-tune | Add or improve a narrow capability | Data lineage, regularization, forgetting tests |
| Adapter train | Cheap specialization on a stable base | Base digest, adapter rank, interference tests |
| Distill | Compress a teacher or coalition | Teacher quality, synthetic-data review, coverage |
| Quantize | Reduce memory and latency | Per-slice regression, calibration, hardware profile |
| Prune | Remove low-value parameters or experts | Rare-case tests, sparsity support, rollback |
| Expert split | Separate conflicting niches | Router retraining, load balance, population cap |
| Prompt/system policy variant | Change behavior without weights | Versioned prompt, injection tests, no policy bypass |
| Router threshold mutation | Change selection behavior | Traffic simulation, fairness, exploration reserve |
Code-level operators
Code-level changes require a separate trust path. Examples include AST patching, algorithm replacement, kernel configuration, cache policy, batching, or workflow graph edits. The operator output must pass compilation, tests, static analysis, security scanning, determinism checks, and performance profiling.
Mutation budget
mutation_budget <- {
max_candidates: 24,
max_training_gpu_hours: 40,
max_data_examples: 500000,
allowed_operators: ["adapter_train", "quantize", "router_threshold"],
max_parameter_delta_norm: 0.15,
max_permission_delta: NONE,
max_wall_clock_hours: 12,
maximum_descendant_depth: 8
}Adaptive mutation rates
Mutation strength can respond to evidence: increase exploration after stagnation, decrease it after instability, or target failure clusters. Adaptation itself must be bounded and logged. Do not let a candidate choose its own mutation rate based on access to hidden evaluation results.
Validity before fitness
Many mutations produce invalid artifacts. Run cheap validity gates first: package integrity, load test, contract schema, basic sanity, and resource ceiling. Reserve expensive evaluations for candidates that satisfy these prerequisites.
One family at a time
For scientific clarity, begin with one operator family. If an adapter, quantization, router, and prompt all change at once, attribution is weak. Later factorial designs can test interactions explicitly.
Mutation operator scorecard
Evaluate operators by useful-descendant yield, average improvement, diversity contribution, cost, failure rate, reproducibility, and maintenance burden. An operator that occasionally finds a large win but produces many unsafe artifacts may not be viable in routine operation.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.