Models

Models

Five classes implement CalibrationModel. CLI --model keys are in backticks.

Class	CLI key	Mean function	LOD	LOQ rule	Solver
PiecewiseCF	piecewise_cf (default)	$y=\max(c, ax+b)$	yes	sliding window on a CV grid (<=)	discrete knot search
PiecewiseWLS	piecewise_wls	$y=\max(c, ax+b)$	yes	sliding window on a CV grid (<=)	scipy TRF curve_fit
EmpiricalCV	cv_empirical	none	no (inf)	sliding window on calibration levels (<=)	group CV
OriginalWLS	original_wls	$y=\max(c, ax+b)$	yes	single grid point with CV strictly < threshold	scipy TRF least_squares
OriginalCV	original_cv	none	no (inf)	lowest level with CV <= threshold	group CV

I use PiecewiseCF for new curves. I use OriginalWLS or OriginalCV when I need the published script's rule, including its LOQ definition. EmpiricalCV is the model-free check. PiecewiseWLS is the TRF sibling of CF, kept so the two solvers can be compared on the same data.

Important

"Three consecutive passing points" is not the same object in every class. Piecewise bootstrap and delta evaluate a linspace grid from LOD to $\max(x)$. EmpiricalCV evaluates unique positive concentrations. OriginalWLS does not use a window at all.

These numbers are not an agency LLOQ. ICH M10 / FDA bioanalytical text is about QC accuracy and precision at LLOQ, not this estimator.