Foundations Introductory 2 minute read Updated 2026-06-29 UTC

Foundations

Beginner-friendly foundations for model breeding as adaptive ecology: five pillars, core loop, vocabulary, lineage, local-first execution, and positive engineering practice.

Research statusSource-backed synthesis Publication statePublished Reviewed byMichael Kappel Source reports7

Model breeding as adaptive ecology

A model-breeding system is not one model. It is a population of variants, specialists, adapters, routes, evaluations, and records. The useful unit is the ecology: the router that chooses, the specialists that work, the evaluator that measures, the lineage DAG that remembers, and the release packet that explains why a descendant earned a place.

The five pillars

Start with The Five Pillars of Model Breeding: compounding, local-first, frugal, generative, and mutualist. These pillars turn the metaphor into practical design questions.

Ecology vocabulary

TermShort definition
PopulationThe active and archived set of model artifacts for an ecology.
VariantA candidate artifact that differs from a parent.
DescendantA variant with recorded parentage and operator history.
ParentAn artifact or recipe used to create a descendant.
ChampionThe current best default for a niche.
SpecialistA narrow high-value model or adapter.
ChallengerA candidate collecting evidence.
AdapterA compact skill overlay such as LoRA.
RouteA policy decision that chooses a specialist, stack, cascade, ensemble, no-op, or review path.
JudgeAn independent evaluator that produces evidence.
Fitness vectorA multi-dimensional measurement of usefulness.
Lineage DAGParent-child graph of artifacts, recipes, and release decisions.
Evidence packetReviewable proof bundle for a candidate.
No-opDecision to keep the current state because change does not add value.
RetirementRemoving a model from active use while preserving its record.
Mutation budgetBounds on how much a candidate may change.
Diversity preservationKeeping useful coverage instead of forcing one winner.
Local-first runtimeA path to run useful work in the browser, on a device, or on controlled infrastructure.

From metaphor to engineering practice

“Breeding” is an engineering metaphor for controlled generation, comparison, and reuse of model variants. It does not require anthropomorphic claims. The practical work is clear: define a niche, create bounded descendants, measure fitness, preserve lineage, release with evidence, and keep useful diversity.

Foundation guides

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.