Thesis and axioms

Core thesis

A model-breeding system should treat population composition as the primary adaptive object. Individual models matter, but the controlled ecology matters more: which specialists exist, which are active, which can substitute for one another, which are archived as stepping stones, and which are retired because their cost exceeds their marginal value.

This thesis converts the research archive into an engineering doctrine: do not ask only whether a candidate is better than a champion. Ask whether the candidate improves the governed ecology after latency, memory, energy, risk, provenance, operational complexity, and human oversight costs are counted.

Axiom 1: resources are finite

Every model consumes scarce resources: memory, storage, evaluation time, human review attention, carbon, latency budget, security surface, and cognitive load. A system that cannot say no will eventually bloat. The no-op action is therefore not a failure. It is the evidence that the ecology can refuse uneconomic growth.

Axiom 2: contracts precede descendants

A descendant is not useful until it satisfies a contract. A contract defines input types, output types, permitted tools, abstention behavior, uncertainty reporting, resource budgets, and escalation paths. Without contracts, the system is a pile of artifacts rather than a population.

Axiom 3: the evaluator is outside the candidate

Candidates may propose improvements, but they do not control the fitness function that promotes them. This is the boundary between adaptation and self-referential metric capture. The evaluator can evolve through a separate code-breeding process, but candidate model parameters do not get permission to rewrite the scoring rules that judge them.

Axiom 4: lineage is memory

A model ecology remembers through immutable descendant records. Each useful artifact has parents, operator history, training data lineage, evaluation evidence, and a rollback target. If parentage is missing, the ecology cannot learn from successes or failures.

Axiom 5: diversity is a reserve, not decoration

Diversity protects the system from sudden environment shifts. A specialist that is not the current champion may still be worth keeping if it covers a rare task, a low-resource runtime, a different jurisdiction, or an adversarial slice where the champion fails.

Axiom 6: persistence is earned

A module remains active only while it contributes. Long-lived artifacts must repeatedly justify themselves. This applies to models, routers, evaluators, documentation, and even theoretical claims.

FUNCTION accept_axioms(package, ecology, policy)
    REQUIRE package.contract EXISTS
    REQUIRE package.lineage.parents IS NOT EMPTY OR package.origin == "seed"
    REQUIRE package.evaluation.used_external_fitness_function
    REQUIRE package.lifecycle.rollback_target EXISTS

    IF package.costs.total > policy.maximum_unpaid_cost
        RETURN REJECT("Cost exceeds allowed unpaid burden")
    END IF

    RETURN ACCEPT("Package can enter candidate pool")
END FUNCTION

Practical implication

The system is evolutionary only after it can record parentage, generate controlled variation, select under pressure, and delete or retire what does not pay. Before that, it is merely an ordinary model-serving stack with a biological metaphor attached.