Models

A model file is a Gamut file with no runs. It holds one or more models fitted to a study, and declares what they take and what they give: the inputs as parameters, whose bounds are the range the model was fitted on, and labels; the results as scalars and fields, each naming the model and the output that computes it; the mesh its fields live on; and the study's constants.

$ gamut info lift_model.gamut
lift_model.gamut: model file
  parameters  alpha [degree] 0 to 10
  results     cl [1]
  model       fit: example.polynomial reading alpha
  trained on  ...
  file        ... KB

That is the quadratic fitted in the models example, recording the study it was trained on by that file's id.

A model is numbers, not code

A model has a type, naming the software that runs it; its inputs; and its own numbers, stored as plain arrays and values. The format standardizes the interface, never the internals: only the software that owns a type reads its numbers, and a file never holds executable code. Opening a model file from someone else runs nothing.

A type is registered with the library as three operations: load, from the stored numbers to a working model; call, from a table of inputs to the outputs asked for, with their standard deviations, quantiles and derivatives when asked; and save, back to numbers. Python and MATLAB register types as classes.

Evaluating

Evaluating a model file on a table of inputs gives a data file with a row per input: the inputs as parameters and labels, the constants, the mesh, and every result filled by its model. It is an ordinary study, read with the same calls as any other, and it records the model file it came from.

An input outside the range a model was fitted on is evaluated, with a warning (W4); refusing it is the caller's choice. The answer keeps the model's bounds on its parameters, so outside can say which rows lie outside them and by how much.

The built-in model

The type nearest returns, for each input row, the outputs of the closest stored run. It is a baseline ("the closest run we have") and the model every language can evaluate with no other software: its distance is defined to the order of each addition, so every implementation gives the same answer.

Running a model of another type needs the software that owns it, and whatever licence that software requires; licensing is not part of the format.