Start here

Direct answer

ModelBreeder.com teaches how to build adaptive AI systems from governed populations of small, specialized, replaceable models. The site emphasizes source-backed engineering patterns: model breeding, code breeding, teleodynamic resource control, independent evaluation, lineage, reversible releases, and human-reviewed support boundaries.

ModelBreeder.com treats model breeding as a controlled engineering discipline: create bounded descendants, evaluate them with independent evidence, preserve useful diversity, and release only variants that improve the whole system after cost and risk are counted.

The site does not assume that one model should continuously rewrite itself. The safer and more operationally useful pattern is a population of immutable artifacts managed by an external control plane. Models can propose descendants, but they cannot redefine their own acceptance criteria, expand their privileges, or remove the rollback path.

Recommended learning path

1. Read What model breeding means to establish scope.
2. Review the research maturity map so established methods are not confused with proposed terminology or speculative scenarios.
3. Learn both Four-F frameworks: Fast, Flexible, Frugal, Federated and Feed, Fork, Fight, Flee.
4. Understand the two-timescale teleodynamic controller.
5. Study the reference architecture, then its contracts, registry, router, evaluator, ledger, and lineage components.
6. Walk through the core evolutionary loop and the mutation, recombination, diversity, and benchmarking guides.
7. Build the minimal breeder before attempting a distributed or continually adapting system.
8. Apply the safety invariants and release pipeline before any production traffic.
9. Select a concrete system blueprint and adapt it to a measured workload.
10. Use the planning tools and reference library during architecture and release reviews.

The five operating principles

Suggested first implementation

Start with three specialists, not thirty: a router, two narrow models, and an independent evaluator. Keep all artifacts immutable. Store manifests and evaluation cards in a file-backed registry. Run descendant generation offline. Promote through shadow and canary stages. This is enough to test the central hypothesis without creating a complicated multi-agent environment.

PROCEDURE first_model_breeder_experiment(requests, champion, specialist_b)
    registry <- LOAD_SIGNED_REGISTRY()
    frozen_suite <- LOAD_FROZEN_EVALUATION_SUITE()

    FOR each request IN requests
        route <- ROUTER_SELECT(request, [champion, specialist_b])
        answer <- RUN_ISOLATED(route.model, request)
        RECORD_TRACE(request, route, answer)
    END FOR

    candidate <- DISTILL_OR_ADAPT(champion, recent_failure_cases)
    evidence <- EVALUATE(candidate, frozen_suite)

    IF evidence.hard_invariants_pass AND evidence.net_viability > promotion_threshold
        DEPLOY_TO_SHADOW(candidate)
    ELSE
        ARCHIVE_OR_RETIRE(candidate)
    END IF
END PROCEDURE

How to use the source reports

The research library preserves every supplied report in original Markdown. The curated guides reorganize that material and add maturity labels, implementation boundaries, and safety framing. Raw reports are useful for tracing ideas and references; curated pages are the preferred operational documentation.

Expanded theory path

Version 1.5.0 adds a dedicated Benefits section. Read it after the foundations and theory if you want the constructive spine: Capability compounding, Local sovereignty, The adapter economy, Model merging upside, and Positive mutualism.