Build a minimal model breeder

Scope

The minimal breeder is intentionally small: one capability, one champion, one mutation operator, one independent evaluator, one file-backed registry, and a shadow release. It teaches the core control loop without distributed agents or online self-modification.

Components

minimal-breeder/
  registry/
    artifacts/
    aliases/
    audit/
  contracts/
  suites/
  datasets/
    manifests/
  experiments/
  runtime/
  tools/
    build_candidate
    evaluate_candidate
    promote_shadow
    rollback

Workflow

1. Package the current model as champion.
2. Freeze a contract and evaluation suite.
3. Select a narrow failure cluster from approved evidence.
4. Create up to five candidate descendants with one operator.
5. Run validity, task, robustness, safety, and resource gates.
6. Compare with champion and no-op.
7. Move the best candidate to shadow only if the margin is material.
8. Record the decision and archive or retire the remaining candidates.

PROCEDURE minimal_cycle(config)
    champion <- REGISTRY_RESOLVE(config.champion_alias)
    experiment <- LOAD_PREREGISTERED_EXPERIMENT(config.experiment_id)

    candidates <- BUILD_CANDIDATES(
        parent = champion,
        operator = experiment.operator,
        configurations = experiment.operator_configs
    )

    results <- []
    FOR each candidate IN candidates
        ASSERT PACKAGE_VALID(candidate)
        evidence <- EVALUATE(candidate, champion, experiment.suites)
        APPEND results, evidence
    END FOR

    best <- SELECT_MATERIAL_WINNER(results, experiment.thresholds)

    IF best EXISTS
        PROMOTE_ALIAS("shadow", best.artifact_id, approval = experiment.approval)
    ELSE
        RECORD_NO_OP(experiment.id, results)
    END IF
END PROCEDURE

No database design

Store canonical JSON manifests and JSONL audit events. Artifact directories are immutable and content-addressed. Aliases are small atomic pointer files. Generated indexes can be rebuilt at any time.

What to postpone

Do not add learned routing, multi-agent debate, online training, federated updates, automated code mutation, external tools, or a general-purpose objective. Add complexity only when the minimal loop exposes a real limitation.

Success criteria

The pilot succeeds if it can reproduce an experiment, reject a candidate, promote to shadow, roll back, and explain every decision. A large performance gain is useful but not required; operational learning is the main deliverable.