Every run, one chart.

Cook is the whole way in, measured end to end: process start, model load, context create, and the cold prefill that reads the document. Reload is the whole way back to that same context: process start, model load, context create, and the import of the saved state. Fit is an observed lower bound from one run on one machine class, not a maximum. Qualified at depth answers the slider position itself: qualified when at or under an observed passing rung, failed when at or past an observed failure, unresolved between the two, or not measured when no rung has been observed for this row. Speeds at a detent are measured; between detents they are interpolated log linear between observed rungs and labelled so. The saved-context and memory-estimate rate varies with context (it is a property of the model, not a fixed number), so both interpolate the per-token rate from the row's density curve: exact at a measured point, interpolated strictly between two, and not observed outside the measured range, never extrapolated.

Models

Models, most active first.

Each model keeps its own page of runs. The order follows recent publishing activity.