Model Field Guide, for looking past the leaderboard

Field notes

The guide frames evaluation as a sequence of task definition, evidence checking, and boundary testing rather than as a single model-ranking exercise. Its diagnostic material names separate failure layers—including retrieval, OCR, schemas, tools, permissions, and interfaces—while the accepted-result calculator keeps calculations in the browser using costs, review time, retries, and acceptance rate. A historical Stanford CRFM diagram is offered as context, with an explicit warning that it is not a benchmark result.