The scorio/sinf helpers take posterior means and standard deviations and return stopping or allocation decisions. Your code runs the trials and updates the estimates. The helpers use normal approximations, so their decisions depend on that approximation and the uncertainty estimates you supply. See the API reference for the available functions.

Stop when an interval is narrow enough

import { bayes } from "scorio/eval";
import { ciFromMuSigma, shouldStop } from "scorio/sinf";

const [mean, std] = bayes([[0, 1, 1, 0, 1], [1, 1, 0, 1, 1]]);
console.log(ciFromMuSigma(mean, std, { confidence: 0.95, clip: [0, 1] }));
console.log(shouldStop(std, { confidence: 0.95, maxHalfWidth: 0.05 }));
Copy

Pass exactly one of maxHalfWidth or maxCiWidth to shouldStop. The latter is the full interval width. Both use the requested confidence level, which defaults to 0.95.

Check the leading model

import { shouldStopTop1, suggestNextAllocation } from "scorio/sinf";

const means = [0.81, 0.79, 0.55];
const stds = [0.03, 0.03, 0.02];
const decision = shouldStopTop1(means, stds, { confidence: 0.95 });
console.log(decision); // => {"stop": false, "leader": 0, "ambiguous": [1]}

console.log(suggestNextAllocation(means, stds)); // => {"leader": 0, "competitor": 1}
Copy

leader and the entries in ambiguous are zero-based model indices. The default "ci_overlap" method stops when the leader's lower interval endpoint exceeds every competitor's upper endpoint. With { method: "zscore" }, it instead compares pairwise ordering probabilities to the confidence threshold.

suggestNextAllocation returns the leader and the competitor with the smallest separation under the same method. It requires at least two models and does not assign a trial budget or start sampling.

rankingConfidence returns { rho, z } for two means and standard deviations. The calculation combines their variances as independent estimates; it does not take a cross-model covariance.