Evolution Lab Intermediate 2 minute read Updated 2026-06-26 UTC

Evolutionary operators catalog

A practical catalog of safe model-breeding and code-breeding operators with evaluation requirements.

Research statusEngineering synthesis Publication statePublished Reviewed byMichael Kappel Source reports3

Operator categories

A breeding lab needs a constrained operator catalog. Operators define what kinds of changes are admissible. Without a catalog, every candidate becomes a special case, and governance becomes informal.

OperatorDomainOutputPrimary evaluator
fine-tunemodeltask specialisttask + drift suite
adapter trainmodelLoRA or adaptercompatibility + task suite
distillmodelsmaller studentteacher agreement + external holdout
quantizemodellower precision artifactaccuracy and latency suite
prunemodelsparse artifactrare-capability regression
merge adaptersmodelfused specialistinterference tests
split expertmodel architecturenarrower expertsrouting and load balance
route policy editcode/confignew selectortraffic replay and fairness
cache policy editcodefaster runtimecorrectness and latency
evaluator editgovernance/codenew test behaviorhuman review required

Operator cards

Each operator should have an operator card with required inputs, forbidden inputs, resource limits, reproducibility steps, and scorecard requirements.

pseudocode
operator_card = {
    name: "distill_specialist",
    allowed_parent_states: ["champion", "approved-teacher"],
    required_inputs: ["teacher_id", "training_case_manifest", "student_base_id"],
    forbidden_inputs: ["hidden_holdout_labels", "production_secrets"],
    output_state: "candidate",
    max_gpu_hours: 12,
    required_evaluations: ["task", "calibration", "safety", "latency", "teacher_error_inheritance"]
}

Safe defaults

Start with distillation, adapter training, and quantization. Delay architecture search and evaluator edits until the registry, scorecards, and rollback process are reliable. Delay any autonomous code breeding until the code pipeline has strong tests and sandboxing.

The operator approval test

Ask three questions before adding an operator:

  1. Can we reproduce the output?
  2. Can we detect the most likely regressions?
  3. Can we rollback the result without affecting unrelated capabilities?

When the answer to any question is no, the operator belongs in research mode, not release mode.

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.