measure before you adopt

Every claim about a model or a tool is written down with one of four labels: verified (read at the source or measured here), estimated (inferred from something verified), hypothesis, unknown. Marketing sentences are quoted as claims, not repeated as facts.

A model gets a number on our own task, against realised outcomes, before it gets a place in a product. The number sits next to the cheapest baselines that could have produced it, and next to the cost. Probabilities are judged with a proper scoring rule, not by how confident they sound.

Every outside service sits behind an interface we wrote, so a swap is credentials and one file, and the lab builds the small in-house version of anything we depend on. That is how sam started.