gpd-parameter-sweep

Define the parameter name, range as start:end:steps, and observable, then the sweep generates one plan per grid point. Points execute in parallel waves, and results aggregate into a sweep-results.json file plus a markdown summary table.

Use linear spacing for smooth observables and logarithmic spacing for parameters spanning orders of magnitude like couplings or masses. Phase diagrams need 20-50 points per dimension, while scaling law fits need 10-20 log-spaced points.

Adaptive refinement computes numerical first and second derivatives of the results and flags regions where change exceeds three times the median, where monotonicity breaks, or where data gaps exist. New points are added in those regions, executed, and merged into the dataset.

Yes, specify two parameters with their own ranges to generate a Cartesian product grid, such as a 10x10 temperature-coupling grid. Each pair is an independent computation executed in parallel waves.

A failed point is recorded in its result JSON with status agent_failed and a null observable, and the sweep continues with remaining points. Failed points are documented in the SWEEP-SUMMARY.md under a Failed Points section.

NaN or Inf results typically come from division by zero, overflow, or underflow at special parameter values and are excluded from feature analysis. Identical values across all points mean either the parameter does not affect the observable or the parameter is not being set correctly.