ControlArchitectural inferencev1.10.0
Reproduction Rate and Resource Limits
Evidence levelArchitectural inference
A model ecology becomes harder to govern when candidate generation outruns review. Rate and resource limits convert an unbounded search into an inspectable process.
What to bound
Bound candidate count, adapter stack depth, total active adapters, merge attempts, memory writes, synthetic-data retention, evaluator invocations, tool calls, network access, canary traffic, and promotion frequency.
Why limits are not enough
A small number of highly influential candidates can still create risk. Limits reduce blast radius; they do not replace composition-aware evaluation.
Healthy limit design
Limits should be explicit, externally enforced, signed, reviewed in UTC release history, and tied to no-op outcomes. Exceeding a limit should trigger a pause, not an automatic search for a workaround.