Model Dossiers

Which model is actually good at which job? Not on a leaderboard, but in your studio, on your work. A model dossier is one model's track record, built from the receipts your own projects leave behind.

A track record, not a benchmark

Every model you run leaves receipts. Did it deliver a usable result, or did it truncate, time out or fail? Did its output pass the quality gate? How long did a finished task take, and how many tokens did it spend? The dossier gathers those receipts model by model, so you can see what each one is really like to work with.

Strengths only when the evidence says so

A dossier does not guess. It names a strength or a weakness only when the numbers separate a model from its peers measured the same way: at least ten observations, a material gap, and a margin of error that clears the pack. Every figure shows its sample size, its range, the conditions it was measured under and the receipt it came from. What a model's profile only declares is labelled as declared, never passed off as a measurement.

“Which of my models is most reliable for chapter drafts, and what is that based on?”

You choose, with the record in front of you

The dossier informs your choices; it never makes them. Use it to decide which model to put on a kind of task, and which two or three contenders to bring into a Workshop run. Nothing in a dossier switches a model, reroutes a task or changes your setup on its own.

Ask in plain language

Open Model Dossiers in the Command Center, or ask the assistant: "What is this model good at?" or "Which model should I use for coding?" The answer comes back with the same evidence and the same receipts.

Included, and it grows with your work

Model Dossiers is part of the core studio, in every package. Your dossiers are built from your own work: a new studio starts with little evidence, and the record fills in as your projects run. It is the book that Eval keeps on the workers.