Theory Advanced 2 minute read Updated 2026-06-26 UTC

Population dynamics

Operational rules for carrying capacity, turnover, archives, starvation, extinction, and active population health.

Research statusConceptual synthesis Publication statePublished Reviewed byMichael Kappel Source reports3

The active population is not the archive

A model ecology has several populations. The active population serves traffic. The candidate pool is under evaluation. The quality-diversity archive preserves useful alternatives and stepping stones. The cold archive preserves history for reproducibility. Mixing these populations causes governance errors.

Carrying capacity

Carrying capacity is the maximum active population that the system can maintain without degrading its viability. It depends on memory, latency, ownership, review capacity, security surface, and router complexity.

pseudocode
FUNCTION carrying_capacity(ledger, policy)
    capacity <- MIN(
        FLOOR(ledger.memory_available / policy.avg_module_memory),
        FLOOR(ledger.review_hours_available / policy.review_hours_per_module),
        FLOOR(policy.max_router_complexity / policy.complexity_per_module),
        policy.absolute_active_module_limit
    )
    RETURN MAX(0, capacity)
END FUNCTION

Population health indicators

IndicatorHealthy signWarning sign
Turnoverweak modules are retired regularlyold modules persist without evidence
Coverageimportant niches have coveragerare slices have no specialist
Starvationuseful specialists are routed occasionallyrouter never selects some active modules
Correlationerrors are partially decorrelatedmany modules fail identically
Age mixmature champions plus challengersonly old models or only unproven candidates
Archive reusearchived elites seed new experimentsarchive is a write-only graveyard

Starvation and overfeeding

A model can starve when a router never selects it, preventing useful evidence from accumulating. A model can be overfed when the router sends too much traffic to it because it is cheap or overconfident. Both distort selection.

pseudocode
FUNCTION rebalance_traffic(population, routing_stats, policy)
    FOR module IN population.active
        IF routing_stats[module].traffic_share < policy.min_evidence_share
            ASSIGN_SHADOW_EVALUATION(module)
        END IF
        IF routing_stats[module].traffic_share > policy.max_share AND module.not_required
            APPLY_ROUTER_PENALTY(module)
        END IF
    END FOR
END FUNCTION

Extinction is sometimes correct

An entire model family may be retired if its runtime is obsolete, its license becomes unacceptable, its tokenizer causes excessive incompatibility, or newer families dominate every useful niche. Preserve history, but do not keep unfit families active for sentimental reasons.

The archive protects future optionality

A cold module may become useful again after a workload shift, hardware change, or new recombination method. Archive enough metadata to recover it safely, but do not let archived modules consume production attention.

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.