Bayesian multi-armed bandits for continuous prompt experimentation: Thompson sampling routes traffic to the best prompt variant and a stopping rule promotes a winner without a fixed-N A/B test. Zero dependencies, TypeScript-first.

Exact Bayesian trust and reputation scoring for tools, MCP servers, skills, and agents. Zero-dependency, TypeScript-first: closed-form Beta-Bernoulli posteriors, calibrated credible intervals, Thompson-sampling routing, time-decay, and a tamper-evident audit trail.

Thompson-sampling Bayesian router that picks the right LLM per query across quality, cost, and latency. Zero-dependency, TypeScript-first, with exact conjugate posteriors, safe exploration, and bounded regret.