[PUBLISHED ON NPMJS] @takk/bayesroute@1.0.0

STATUS: PUBLISHED ON NPMJS. This version was published to the npm registry on 2026-06-20T15:56:59Z with provenance attestation. View on npm: https://www.npmjs.com/package/@takk/bayesroute/v/1.0.0

STATUS: REVIEW REQUIRED, NOT YET ON NPMJS. This GitHub Release was created by the release.yml workflow. The Creator must review the contents (tag, changelog, attached commit, pack-smoke result in the workflow logs) and then explicitly run npm-publish.yml to publish this version to the npm registry.

[1.0.0] - 2026-06-20T14:57:44Z

Initial stable release. BayesRoute is a zero-runtime-dependency, TypeScript-first Thompson-sampling Bayesian router that picks the right model for every query across quality, cost, and latency for Massive Intelligence (IM) products and agents. Model selection becomes a calibrated probabilistic decision with bounded regret, not a static rule or a greedy top-model heuristic.

Added

Core router

• createBayesRoute({ models, ... }) facade wiring the query classifier, posterior store, Thompson sampler, multi-objective utility, safe-exploration guard, optional time-decay, and the tamper-evident audit log into one object.
• route(features) classifies the domain, Thompson-samples the (model, domain) arms, and applies the safe-exploration guard.
• run(features, call) one-call path: route, call the chosen model, measure latency with the configured clock, and fold the outcome back in a single await.
• observe, observeDecision, arm, summary, rank, models, domains, snapshot, load, auditLog, seal, and verify.

Conjugate posteriors

• ./beta: the Beta-Bernoulli quality posterior, exact moments, density, CDF via the regularized incomplete beta, quantile, equal-tailed credible interval, conjugate update, and seeded sampling.
• ./gamma: the Gamma-Exponential latency and cost posteriors, prior-from-mean, expected-value mean, conjugate update, seeded sampling, and exact value credible intervals via the regularized incomplete gamma.
• Internal special functions: lgamma (Lanczos), the regularized incomplete beta and gamma and both quantiles, exact to numerical precision.

Decision layers

• ./arms: the per (model, domain) posterior bundle, init, observe, and calibrated summary.
• ./sampler: the Thompson sampler and the greedy incumbent.
• ./utility: the multi-objective expected utility with configurable weights and reference scales.
• ./classify: default, static, and keyword query classifiers, plus support for any custom function.
• ./outcome: scorer-agnostic outcome constructors plus heuristic evaluators (lengthEvaluator, regexEvaluator, jsonValidEvaluator, combineAll, combineAny, evaluate) and an async LLM-as-judge type.
• ./policy: the safe-exploration guard that bounds how far a single decision may stray from the greedy incumbent.
• ./decay: time-decay of evidence for non-stationary models and silent provider deprecations.

Persistence and audit

• ./store: the PosteriorStore interface, an in-memory store, and JSON snapshot encode, decode, and load with full field validation.
• ./node: createFileStore, atomic file persistence, the only entry that touches a Node built-in.
• ./edge: the node-free core re-exported for edge runtimes and the browser.
• ./audit: an append-only SHA-256 hash-chain audit log built on the Web Crypto API, with seal and verify for tamper evidence.

CLI

• Binary bayesroute with route, observe, rank, inspect, verify, help, and --version. JSON output; sysexits-style exit codes (0 ok, 1 error, 64 usage, 65 data, 66 no input).

Distribution and quality

• Dual ESM and CJS with per-subpath type definitions; fourteen subpath exports plus the CLI; a files allowlist; engines node >= 20.
• 511 tests across 22 suites; coverage statements 98%, branches 96%, functions 99%, lines 98%.
• TypeScript 6 in maximum strict mode, Biome 2, publint and attw clean across all subpaths, size-limit budgets (the core is about 5 kB brotli), a distribution smoke test that exercises the compiled ESM and CJS artifacts and the compiled CLI, validated on the CI matrix across Node 20, 22, and 24.
• Release publishing uses provenance (SLSA attestation) through GitHub Actions.

Documentation and examples

• README, SPEC, SECURITY, PRIVACY, NOTICE, CLA, CODE_OF_CONDUCT, the issue and pull-request templates, and the GitHub Pages site written for BayesRoute.
• Runnable examples: basic routing, the run one-liner, the built-in quality evaluators, a deterministic savings benchmark against always-frontier and always-cheapest baselines, and a Vercel AI SDK and Mastra integration guide.
• A real-data benchmark (benchmarks/) on the public RouterBench dataset (36,497 real queries across 11 models): replaying the stream through BayesRoute, it matches the spend of always-using GPT-3.5 at higher accuracy (69.6% versus 66.6%) and reaches 73.1% accuracy at 13% of GPT-4's cost, learning per-domain online with no hand-written rules. The benchmark and its dataset are development-only and excluded from the published tarball.