How the courses are made.

Runs Clean publishes engineering courses on the production layer of AI infrastructure: the protocols, gateways, auth and operations that sit between a model and a system people depend on. The name is what a developer says when a build comes back green. It's also the bar.

The rules every course follows

Labs run before anything is narrated. A lesson is written against a lab that already passes. If the lab changes, the lesson is checked again. Every lab runs locally in Docker with images pinned by digest, and none of them asks for an account or an API key.

Claims are written down with their source. Each fact a lesson relies on is recorded with the section of the specification it comes from and the date it was checked. MCP Production alone has 224 of those records.

Tests are checked for teeth. Every lab ships with checks that break the code on purpose. A passing suite that doesn't notice a deliberate defect isn't evidence of anything.

Gates keep running after the tests pass. During an audit before launch, a gate that checks which JSON-RPC error codes a lab is allowed to emit found a gateway answering its refusals with -32000. The revision tells new servers not to use that range. Every test had passed, because none of them asked which code a refusal carried. The lab was fixed, and the gate now runs on every lab of every course.

Courses are dated. Specifications move. Each course says which revision it covers and when its sources were last checked, so you can tell what may have changed since.

What we don't do

There's no invented instructor. No made-up name, no stock headshot, no résumé that falls apart when someone looks closely. The courses are published as Runs Clean.

What you're paying for is the labs, the tests, and the checks that hold every lesson to what the spec actually says.

Get in touch

Found something wrong in a lesson, or a spec change we missed? That's the most useful email we get: [email protected].