OriginalCV

OriginalCV

CLI: --model original_cv. Class: loqculate.compat.cv.OriginalCV. Registered on MODEL_REGISTRY when you import loqculate.

Port of old/loq_by_cv.py calculate_LOQ_byCV(). No regression, no LOD, no sliding window.

Fit

Groups by unique concentration. CV is Bessel-corrected std / mean (vectorized_cv_stats), nan when count $< 2$ or mean is 0.

LOQ is computed during fit():

$$\mathrm{LOQ} = \min{ C > 0 : \mathrm{CV}(C) \le \tau }$$

$\tau$ is the constructor cv_thresh (default 0.20). Non-strict <=, matching the original <= 0.2 filter. If no level qualifies, $\infty$.

loq(cv_thresh) after fit ignores the argument and returns the stored value. CLI --cv_thresh is therefore unused for this model. --sliding_window and --bootreps are unused.

lod() is always $\infty. supports_lod()` is false.

Predict

predict() raises NotImplementedError. There is no curve.

Compared with EmpiricalCV

Same CV estimator. Different LOQ rule: OriginalCV is single-point <=; EmpiricalCV requires sliding_window consecutive positive levels. comparison_report.ipynb treats OriginalCV FDR as the window=1 case (64-99% across the concentration-density experiment in that notebook).