Core claim
An adaptive model ecology is a population of capability packages managed by a control plane. The population includes champions, specialists, challengers, adapters, routes, and archived parents. The control plane keeps evidence, lineage, resource budgets, and release decisions separate from the model artifacts being evaluated.
The ecology improves when useful descendants become reusable parents. This is the heart of model breeding.
Why ecology is the right metaphor
Ecology emphasizes relationships. A specialist matters because it fits a niche. A router matters because it connects requests to the right capability. An evaluator matters because it turns outputs into evidence. A lineage record matters because it lets the next generation inherit useful work.
Practical theory
| Concept | ModelBreeder interpretation |
|---|---|
| Niche | A measurable capability contract. |
| Fitness | Multi-dimensional evidence of usefulness under budget. |
| Diversity | Preserved coverage across meaningful behavior and resource profiles. |
| Carrying capacity | The number of active artifacts the team can maintain well. |
| No-op | A decision to preserve the current ecology when change does not repay cost. |
| Retirement | Removing an active artifact while retaining its learning value. |
FUNCTION ecology_step(ecology)
evidence <- COLLECT_RECENT_EVIDENCE(ecology)
candidates <- PROPOSE_BOUNDED_DESCENDANTS(ecology.parents, evidence)
scored <- SCORE_WITH_FITNESS_VECTOR(candidates)
return SELECT_PORTFOLIO(scored, ecology.carrying_capacity)
END FUNCTIONSource 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.