Core evolutionary loop

End-to-end loop

The core loop converts evidence into controlled structural change. It combines the evolutionary Four Fs with software-release discipline.

PROCEDURE breeding_cycle(ecology, evidence_window, policy)
    trigger <- DETECT_ADAPTATION_TRIGGER(ecology, evidence_window, policy)
    IF trigger.none
        RETURN RECORD_NO_OP("no material trigger")
    END IF

    feed <- CURATE_TRAINING_AND_EVALUATION_INPUTS(trigger, policy)
    experiment <- REGISTER_EXPERIMENT(
        hypothesis = trigger.hypothesis,
        parents = SELECT_ELIGIBLE_PARENTS(ecology, trigger),
        operator_family = SELECT_OPERATOR_FAMILY(trigger),
        budget = policy.experiment_budget
    )

    candidates <- []
    FOR each proposal IN GENERATE_BOUNDED_PROPOSALS(experiment)
        candidate <- BUILD_IN_ISOLATION(proposal)
        package <- PACKAGE_IMMUTABLY(candidate, proposal.lineage)
        APPEND candidates, package
    END FOR

    evidence_cards <- EVALUATE_INDEPENDENTLY(
        candidates = candidates,
        champion = ecology.champion,
        suites = policy.frozen_suites
    )

    decision <- VIABILITY_CONTROLLER_DECIDE(
        evidence_cards,
        ecology.resource_ledger,
        policy
    )

    IF decision.action == "shadow"
        SEND_TO_RELEASE_PIPELINE(decision.candidate)
    ELSE
        ARCHIVE_OR_RETIRE(candidates, decision)
    END IF

    APPEND_CYCLE_AUDIT(experiment, evidence_cards, decision)
END PROCEDURE

Trigger detection

Valid triggers include a persistent error cluster, task-distribution drift, a new hardware target, cost pressure, a revoked dataset, a new compliance boundary, or a declared capability request. Avoid triggering on isolated user dissatisfaction or short-term metric noise.

Parent selection

Choose parents for a reason. A champion provides stability; a specialist provides niche capability; an archived variant may contribute diversity. Record the expected inherited traits and compatibility assumptions.

Candidate generation

Use a fixed candidate count and operator budget. Randomness is reproducible through recorded seeds. Operators can be adaptive across cycles, but not silently changed within one comparison.

Evaluation and selection

Compare against the champion and no-op. Use confidence intervals and slice analysis. A candidate that wins average utility but loses a critical safety or minority slice is rejected or constrained to a narrower contract.

Lifecycle outcomes

• Promote to shadow: evidence supports further production-like observation.
• Archive: useful diversity or research insight without current deployment value.
• Quarantine: policy, security, provenance, or unexplained behavior concern.
• Retire: duplicate, invalid, or uneconomic artifact.
• No-op: no candidate justifies structural change.

Cycle health metrics

Track candidate yield, valid-package rate, evaluation cost, time to evidence, promotion rate, no-op rate, rollback rate, diversity gain, and cumulative operational complexity. A very high promotion rate can indicate weak gates; a zero promotion rate can indicate a broken search space or stale evaluation.