The craft of llm.fit
The interface treats an estimate as an argument, not a decoration. It leads with the fit verdict and decode range, keeps the inputs visible, and puts the arithmetic and evidence close enough to inspect.
Show the estimate and its range
Every result carries p10, p50, and p90 estimates rather than one overconfident number. The artifact also requires a calibration cohort size, evidence URLs, and a verification command with each estimate.
Use exact inputs where they exist
The Hugging Face resolver reads model architecture from config.json and exact artifact bytes from repository files. That avoids pretending a name or a nominal parameter count is enough to predict fit.
Keep provenance visible
Calibration uses public measured runs from the llmfit corpus, localmaxxing, llama.cpp scoreboards, and LocalScore. Formula-only estimates drift. A measured result on the target machine remains the final check.