Cost is a blended rate of input×0.25 + output×0.75 (assumes a 1:3 input:output token mix — adjust PRICE_MIX in the source if your workloads differ). Value = performance score ÷ blended cost. Scores and letter grades are taken as-is from the loaded matrix file; the current build has known scoring caveats (see the v2 build spec). Models with no grade in the selected capability are dimmed and excluded from the plot and the top picks. Disk, runtime, and system RAM are estimates for Ollama models. The memory filter uses recommended total system RAM, not raw model disk size.