ModelBreeder MVP

Goal

The first ModelBreeder MVP should not attempt unrestricted online evolution. It should prove the disciplined loop: package, evaluate, compare, archive, promote, rollback, and retire.

Phase 1: registry and manifests

Create a file-backed registry that stores immutable package manifests, artifact hashes, source reports, scorecards, and lifecycle states. The registry is the memory of the population.

Phase 2: fixed evaluator

Build a frozen evaluator with representative test slices, resource profiling, calibration checks, and hard gates. The evaluator should be outside candidate control.

Phase 3: candidate generation

Start with conservative operators: prompt template variants, deterministic routing policies, adapters, and distilled small specialists. Avoid code self-modification and broad permissions.

Phase 4: shadow and canary

Route copied traffic to candidates without affecting users. Promote only after scorecards, resource budgets, and rollback are ready.

Phase 5: retirement

Add a weekly review that retires or archives weak candidates. A breeding MVP without deletion is just artifact accumulation.

Minimal loop

FUNCTION modelbreeder_mvp_cycle()
    candidates <- GENERATE_LOW_RISK_CANDIDATES()
    FOR candidate IN candidates
        package <- CREATE_PACKAGE(candidate)
        scorecard <- RUN_FIXED_EVALUATOR(package)
        decision <- APPLY_VIABILITY_POLICY(package, scorecard)
        WRITE_DECISION_LEDGER(decision)
    END FOR
    REVIEW_RETIREMENTS()
END FUNCTION

Definition of done

The MVP works when a new descendant can be created, evaluated, compared against a baseline, rejected or archived with evidence, promoted through a canary, and rolled back by digest. It does not need autonomous breeding to be valuable.