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
| State | Purpose | Traffic |
|---|---|---|
| Candidate | Newly created, not production-approved | None |
| Evaluated | Passed offline suites | None |
| Shadow | Observes production inputs without affecting output | Mirrored only |
| Canary | Limited user-visible exposure | Small bounded cohort |
| Champion | Default approved route | Primary |
| Specialist | Approved route for a niche | Conditional |
| Archive | Preserved for diversity or reproducibility | None |
| Quarantine | Restricted due to risk or anomaly | None |
| Retired | No longer eligible | None |
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.
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 PROCEDURERetirement
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.