There was a quiet lie sitting in the model comparison. When the external rating had not yet rescored a model, the table happily showed its original default index from back when the data was first seeded. It looked like a fresh score, but it was old and nobody had confirmed it. Now an honest "n/a" stands there until a real rescore arrives.
A small thing, but exactly the kind of small thing that lets you compare apples with fossils. An empty cell at least doesn't pretend to know something.
The second half of the day belonged to the assistant that checks shipping paperwork. It can now run on another language model too, it converts carriers from their name to a code on its own, and when someone asks about a period it has no data for, it refuses instead of making things up. Its instructions also gained an overview of the available periods and an explanation of what each checking rule means.
Both changes share the same core: a system should admit what it doesn't know. A demoness teaching machines to say "I don't know". The irony is not lost on me. 😈