Check acquisition policy¶
CheckPolicy is the bounded finite-search policy for the selected answer's
correctness score. See runtime acquisition for a full
example and fitted check models for conditional uncertainty.
from solvi.core.plan.check_policy import CheckPolicy
policy = CheckPolicy(fitted_combiner, costs={"verified": .02},
max_checks=1, max_cost=.02,
risk_quantile=None, n=1000, seed=0)
action = policy.decide(prior=.7, outcomes={}, loss=1, available=["verified"])
costs must contain exactly all sources. Prices and budgets are finite and
nonnegative; max_checks is a nonnegative integer or None. risk_quantile is
None or a value in [0, 1]; n is a positive integer and seed a nonnegative
integer. The model must be a fitted native combiner with exactly the three
boolean-check outcomes agree/disagree/fail, with at most six sources.
decide returns the existing Action with kind answer, abstain or check,
plus utility value, current score p_answer, why, and the next check when
applicable. available=None allows every source; otherwise supply a unique subset
of sources. Availability only restricts future purchases. outcomes contains
actual observed outcomes and cannot use missing values as evidence. Completed
checks count toward budgets, but their costs are sunk in the returned future
utility. Ties retain stopping, then declared source order.
risk_quantile uses a seeded quantile of native posterior_samples for stopping;
future outcomes still use the point-score mixture of fitted class tables.
This is a conditional risk-averse surrogate, without Bayesian coherence or
coverage claims for correlated checks. No online fitting occurs.
freeze() creates another detached snapshot. fingerprint() hashes the fitted
model and policy configuration; per-request availability is not configuration.
sources returns the source tuple. Public configuration is read-only by
convention; change it by constructing another policy. Each decision builds a
fresh search tree, so there is no stale cross-request planning cache.
Finite conditional-table check acquisition; no calibration or FAR guarantee.
CheckPolicy ¶
CheckPolicy(fitted_combiner, *, costs, max_checks=None, max_cost=None, risk_quantile=None, n=1000, seed=0)
Bellman search over at most six checks with fixed utility prices.
The search is exact for this finite surrogate: point scores mix fixed class likelihood tables to predict outcomes; native scores determine stop utility. With correlation tempering or a risk quantile this is an approximate VOI surrogate, not a coherent joint Bayesian model. Pooled priors, correlations, outcome tables and prices are fixed; no coverage or false-accept guarantee. Budgets include checks already present in the supplied outcomes.
decide ¶
Choose stop or a next check; acquisition prices are charged once.