Foundations Introductory 2 minute read Updated 2026-06-26 UTC

Terminology and naming

A normalized vocabulary for model breeding, code breeding, code beading, the two Four-F frameworks, and related evolutionary terms.

Research statusEditorial synthesis Publication statePublished Reviewed byMichael Kappel Source reports3

Use separate names for separate control problems

The research corpus contains several overlapping terms. ModelBreeder.com uses the following normalized vocabulary so discussions remain operationally precise.

TermPreferred meaningAvoid implying
Model breedingCreating and selecting parameterized model descendants.Autonomous reproduction or biological agency.
Code breedingControlled variation of algorithms, routing, kernels, workflows, or infrastructure code.Allowing a model to change its own evaluator or security policy.
Code beadingDecomposing work and memory into small, persistent, linked units with acceptance criteria.A recognized universal technical standard.
Model ecologyA managed population of specialists and shared services.A literal ecosystem with intrinsic drives.
Operational 4FsFast, Flexible, Frugal, Federated.The biological Four Fs.
Evolutionary 4FsFeed, Fork, Fight, Flee.Unbounded competition or self-preservation.
Teleodynamic controllerA resource-aware, two-timescale structural control loop.Conscious purpose or independent moral status.
FitnessAn evaluation vector or governed aggregate score.A single accuracy metric.
LineageImmutable parentage and transformation metadata.Genetic identity.
No-opExplicit decision not to change the system.Failure to innovate.

Code breeding versus code generation

Code generation produces a candidate implementation. Code breeding adds population, inheritance, mutation operators, independent evaluation, selection, and lineage. A one-off generated patch is not a breeding system. The distinction matters because population-based search can exploit weak tests more aggressively than ordinary development.

Code beading versus model modularity

Code beading concerns persistent work decomposition and agent memory: small issues, dependencies, acceptance criteria, and handoff state. Model modularity concerns deployable capability packages. They can support each other—an experiment can be decomposed into beads—but they are not the same architecture.

Specialist, expert, module, and bead

Use specialist model for an independently versioned model with a narrow capability. Use expert when referring to a component inside a mixture-of-experts architecture. Use module for a general composable software or model component. Use bead only when emphasizing a small linked unit of work, memory, or capability in the site's conceptual framework.

Fitness, viability, and utility

  • Utility measures task value or business outcome.
  • Fitness is the evidence used to compare candidates under a defined environment.
  • Viability is system-level fitness after subtracting lifecycle costs and enforcing hard constraints.

A candidate can improve utility while reducing viability—for example, a 1% accuracy gain that doubles latency, introduces an unlicensed dataset, and removes rollback compatibility.

pseudocode
viability <- utility_gain
             + robustness_gain
             + diversity_gain
             + coverage_gain
             - memory_cost
             - latency_cost
             - energy_cost
             - risk_cost
             - maintenance_cost

IF any_hard_invariant_fails
    viability <- REJECTED
END IF

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.