@lacspace/postbandit

Learn the best option per platform from real engagement — posting time, format, thumbnail or title variant. A tiny Thompson-sampling multi-armed bandit: each arm keeps a Beta posterior, choose() samples and picks, update() folds in a reward. Deterministic when seeded, persistable to JSON. Designed to plug into @lacspace/postplan.

npm i @lacspace/postbandit

import { createBandit } from "@lacspace/postbandit";

const bandit = createBandit(["09:00", "13:00", "19:00"], { seed: 1 });

const slot = bandit.choose();            // which posting time to try next
// …post at `slot`, later measure engagement (0..1)…
bandit.update(slot, 0.42);               // normalized engagement rate

bandit.best();                            // the current winner (exploit)
bandit.stats();                           // [{ id, alpha, beta, pulls, mean }], best first
const state = bandit.toJSON();            // persist; restore with Bandit.fromJSON(state)

Why

A/B-testing post timing/format/variants by hand never converges. A bandit balances exploring new options against exploiting the winner, and improves every time you feed it a result — no model, no tokens.

API

• createBandit(ids?, { priorAlpha?, priorBeta?, seed? }) — Beta(1,1) priors by default; pass a seed for reproducible sampling (tests/replay).
• choose() → the arm id to try next (Thompson sample). rank() → all arms ordered by one sample each (for an ordered lineup).
• update(id, reward) — reward in [0,1] (booleans map to 1/0). Unknown ids are added on the fly.
• stats() → { id, alpha, beta, pulls, mean }[] (best mean first). best() → top arm id.
• toJSON() / Bandit.fromJSON(state, opts) — persist and restore across runs.
• describe() → machine-readable command schema for an AI "conductor".

Use one bandit per (platform × decision) — e.g. a format bandit whose arms are video|image|carousel|text, fed back from @lacspace/postplan's choices and their engagement.

Licence

Lacspace Free Licence v1.0 — free for personal and commercial use.