Working notes on the methods, the standards behind them, and what we are building.
A parts-count prediction is a sum, and every part it cannot compute still has to contribute something to that sum. When the answer is nothing, the error always runs the same way: optimistic, and invisible. On the four states a part can be in, why an uncomputable part is excluded rather than counted as zero, and the difference between a system that models out at zero and one that modelled nothing at all.
Read it →We would rather publish rarely than fill this with posts nobody needed written. If there is something you want covered — how a standard behaves in a case you keep hitting, or why ReliFrame does something the way it does — say so. That is a better way to choose the next one than guessing.
The documentation covers how the methods work and which standards sit behind each prediction.