Evolution Lab Advanced 2 minute read Updated 2026-06-26 UTC

Population management

Rules for champions, challengers, specialists, archives, duplicates, age, traffic allocation, and retirement.

Research statusConceptual synthesis built from portfolio and lifecycle management Publication statePublished Reviewed byMichael Kappel Source reports2

Population is a budgeted portfolio

Every active model consumes storage, memory, evaluation capacity, operational attention, and attack surface. Population management balances coverage and diversity against these lifecycle costs.

Population states

StatePurposeTraffic
CandidateNewly created, not production-approvedNone
EvaluatedPassed offline suitesNone
ShadowObserves production inputs without affecting outputMirrored only
CanaryLimited user-visible exposureSmall bounded cohort
ChampionDefault approved routePrimary
SpecialistApproved route for a nicheConditional
ArchivePreserved for diversity or reproducibilityNone
QuarantineRestricted due to risk or anomalyNone
RetiredNo longer eligibleNone

Population ceilings

Set maximum active specialists per capability, maximum archive size, maximum descendant depth, and maximum maintenance cost. The controller cannot exceed these ceilings by accumulating individually positive decisions.

Duplicate detection

Detect duplicates using artifact hashes, lineage, behavioral similarity, and highly correlated errors. Two models with different weights but indistinguishable behavior may not justify separate maintenance.

Aging

Age does not make a model bad, but stale evidence does. Revalidate packages when task distribution, policy, runtime, or data assumptions change. Expire eligibility rather than deleting the artifact.

Traffic allocation

Reserve exploration traffic for challengers in low-risk contexts. Protect niche specialists from starvation by maintaining benchmark evidence independent of live selection. Monitor traffic concentration and fallback patterns.

pseudocode
PROCEDURE rebalance_population(population, policy)
    REMOVE_INELIGIBLE(population)
    QUARANTINE_ANOMALOUS(population)
    FIND_AND_MARK_REDUNDANT(population, policy.similarity_threshold)

    FOR each capability IN population.capabilities
        ENSURE_ONE_ROLLBACK_READY_CHAMPION(capability)
        ENSURE_REQUIRED_NICHE_COVERAGE(capability)
        ENFORCE_ACTIVE_MODEL_CEILING(capability)
    END FOR

    SCHEDULE_REVALIDATION_FOR_STALE_EVIDENCE(population)
END PROCEDURE

Retirement

Before retirement, identify dependencies, aliases, cached references, rollback chains, and legal retention requirements. Move traffic to a verified replacement, observe stability, then revoke eligibility. Preserve minimal lineage and audit evidence.

Ecological health indicators

Healthy populations show bounded size, clear niche ownership, low unexplained overlap, stable rollback readiness, measured diversity, and a meaningful no-op rate. Constant churn or one model absorbing all traffic are both reasons to investigate.

Source reports used for this guide

These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.