Most AI you have used only talks. It answers a question, drafts an email, summarizes a report — and a person decides what to do with the output. The stakes are low, because nothing happens until a human moves. The moment you let AI act — deploy a change, edit a live

The easiest thing to build in ed-tech is a machine that writes the paper. The most defensible thing is to refuse. Here is how our verification discipline looks pointed at study tools for real college courses.

Documentation drifts. We pointed our verification discipline inward — a knowledge base as the single source of truth, machine-readable contracts inside the docs, and read-only workers that reconcile live against the contracts and text us the moment they disagree.

One big program that does everything is the least defensible thing you can build. Here is how we took ours apart — a router, one integration per worker, scoped keys, and a second model from a different lab checking the first.

A single model answers with the same confidence whether it is right or wrong. Here is the composable pattern we use to make an answer you can actually check — grounding, an independent reviewer, and a human on the calls that matter.