admissionVerdict
admissionVerdict(
chain,sampleRate,budget):AdmissionVerdict
Defined in: packages/runtime/src/admission.ts:221
Whether chain — an ordered list of Programs, each solved every sample — can be shown to
fit inside budget at sampleRate, given a sustained allowance of
budgetedIterationsPerSample Newton iterations for each nonlinear block.
The allowance is a policy, not the solver’s cap — see the header. A sample that needs more is a spike the overrun policy covers (a bounded solver and a held sample), which is why raising the cap so a legitimate 183-iteration transient can converge no longer makes every chain 16x more expensive to admit.
Still fail-closed on everything it can price: a macro block is charged the host’s own per-step figure, an unpriced model is refused by name, and costs add across slots.
Costs add across slots and nothing here discounts them. There is no shared work between two slots — each holds its own program, its own state and its own solver scratch — so the sum is the honest model, not a conservative one.
Parameters
Section titled “Parameters”readonly Program[]
sampleRate
Section titled “sampleRate”number