// Experiments — 05

Open questions, built small.

Prototypes and proofs of concept — each with a hypothesis and a next step. A prototype means partially built; a concept means not built yet. Labels stay honest.

2 prototypes2 concepts

AGENTIC COMMERCE

Prototype

Updated 2026-10

Agent-Assisted Product-Data QA

An LLM pipeline that reviews a product catalog against a store's own rules — missing attributes, inconsistent naming, weak media, pricing outliers — and returns a reviewable report, not auto-fixes.

H — Hypothesis

Catalog quality is the highest-leverage, least-visible commerce problem. A reviewer that reads the whole catalog the way a merchandiser would — but never tires — can catch the long tail humans miss.

› Next step

Broaden the rule set beyond structural checks to tone and consistency, then measure catch-rate against a hand-audited sample.

DEVELOPER TOOLING

Prototype

Updated 2026-09

Webhook Replay Harness

A local harness that captures live commerce webhooks, stores them, and lets you replay, re-time, and mutate them against a staging environment — so integration failures can be reproduced instead of guessed at.

H — Hypothesis

Most platform integration bugs are timing and ordering bugs. A recording you can scrub through beats logs you can't.

› Next step

Add scenario packs — known failure sequences (duplicate delivery, out-of-order events, replayed stale signatures) as one-command replays.

AGENTIC OPS

Concept

Updated 2026-08

Ops Runbooks as Agent Checklists

Turning operational runbooks (rotate a key, restore a backup, process a refund batch) into structured checklists an agent can execute step-by-step — with confirmation gates before anything consequential.

H — Hypothesis

The value of agents in operations is not autonomy; it's a tireless executor that follows the runbook exactly and never skips the verification step — as long as the gates are real.

› Next step

Encode one existing runbook from the retail-credit platform and run it in dry-execution mode.

INTERFACE

Concept

Updated 2026-07

Checkout Motion Study

A design-only study of checkout micro-states: what should move, what must never move, and how feedback should behave across loading, validation, and error states on the highest-stakes page of a store.

H — Hypothesis

Motion on checkout pages earns its place only when it explains state changes. Everything else is latency with better lighting.

› Next step

Build the state matrix as static prototypes before any code touches a real checkout.

// Engage

Curious is a valid reason to write.

If one of these experiments intersects with a problem you have, that's the fastest way to turn it into something real.

Start a conversation hello@samcommerce.dev