Epigenetic steering and mating kernels

Direct answer

An epigenetic steering genome is a compact parameter vector that modifies behavior without replacing the full model. A mating kernel is a controlled search operator that recombines two compatible parent vectors, adds bounded variation, and produces a candidate that must still pass independent evaluation.

The useful idea is not that models are organisms. The useful idea is that a small, auditable descendant descriptor can sometimes explore a high-value behavior space more cheaply than full fine-tuning or full model retraining.

Design interpretation

The source directive proposes a 20-dimensional continuous vector as an epigenetic steering wheel. ModelBreeder.com treats that as a research pattern:

• the base model is the somatic substrate;
• the steering vector is a compact genotype;
• the projection layer maps the genotype into runtime behavior;
• the mating kernel recombines compatible genotypes;
• the evaluator measures behavior independently;
• promotion requires evidence, lineage, cost accounting, and rollback.

Bounded mating kernel

FUNCTION breed_steering_candidate(parent_a, parent_b, policy, evaluator)
    REQUIRE parent_a.base_family == parent_b.base_family
    REQUIRE parent_a.runtime_contract == parent_b.runtime_contract
    REQUIRE parent_a.tokenizer_digest == parent_b.tokenizer_digest

    distance <- L2_DISTANCE(parent_a.steering_vector, parent_b.steering_vector)
    IF distance < policy.minimum_novelty_distance
        RETURN NO_OP("parents too redundant")
    END IF
    IF distance > policy.maximum_compatibility_distance
        RETURN DISTILLATION_REQUIRED("parents too far for direct recombination")
    END IF

    child_vector <- WEIGHTED_CROSSOVER(
        parent_a.steering_vector,
        parent_b.steering_vector,
        weights = RECENT_SUCCESS_WEIGHTS(parent_a, parent_b)
    )
    child_vector <- ADD_BOUNDED_NOISE(child_vector, policy.mutation_sigma)
    child_vector <- CLIP_TO_BUDGET(child_vector, policy.vector_norm_budget)

    child <- PACKAGE_CANDIDATE(
        base = parent_a.base_family,
        vector = child_vector,
        parents = [parent_a.id, parent_b.id],
        operator = "bounded_steering_kernel"
    )

    result <- evaluator.RUN_HELD_OUT_SUITES(child)
    IF result.viability_margin <= policy.minimum_margin
        RETURN NO_OP("candidate did not repay cost")
    END IF

    RETURN HOLD_FOR_SHADOW(child, result)
END FUNCTION

Why this belongs in the theory layer

This pattern extends the current ModelBreeder vocabulary:

Compatibility rules

Direct steering-vector breeding should require shared base model family, matching tokenizer identity, matching projection layer, matching runtime contract, compatible training/evaluation suites, and stable artifact hashes. If those conditions fail, use distillation, adapter training, routing, or ensemble composition instead of vector recombination.

What is not claimed

This page does not claim that a 20-dimensional vector is universally sufficient, that hybrid vigor is guaranteed, that evolution should be unconstrained, or that candidates should control their own evaluation. The theory becomes engineering only after it is measurable, reproducible, bounded, and reversible.