summary_table()$converged now means the selected model actually fitted. It was !is.na(selection$best_model_name), and the eligibility fallback in curveRcore::select_best_eligible() always returns a name – so a plate on which no nls() succeeded was still reported as converged. This was masked until now by a crash (below) that made the whole per-plate call fail and record the curve as failed for the wrong reason.
New additive fallback column in summary_table(), recording whether the selection came from the eligibility fallback. “Fitted cleanly” and “fitted but passed no gate” were previously conflated; on real plates the latter is common.
Fixed three NULL-fit dereferences.extract_best_parameters() guarded the model name but not the fit and hit summary(NULL)$coefficients ($ operator is invalid for atomic vectors); the grid fallback and the sample path in fit_calibration() had the same shape and reached vcov(NULL). All three now guard the fit, matching the check step 6a already performed.
Validated across 1,706 plates / 119 antigen units in four studies: no convergence count changed, and the 3-level case that per-plate NLS genuinely cannot fit is now reported as clean non-convergence rather than a caught crash.
curveRfreq 0.4.2
Created a new pcov_gate_class and changed the basis for pcov_pass classifcations.
curveRfreq 0.4.1
Fix: predict_samples_freq() now gates pcov_pass on pcov_threshold (not cv_x_max), matching predict_grid_freq(). Previously frequentist test samples with pcov between pcov_threshold and cv_x_max were incorrectly marked pcov_pass = TRUE.
curveRfreq 0.4.0 (2026-07-29)
Lockstep version bump — no functional changes. Released so the curveR stack shares a version and the worker image can pin curveRfreq@v0.4.0 for reproducible builds. Rebuilt/verified against curveRcore 0.4.0.
curveRfreq 0.1.0
Initial release.
fit_calibration_freq() — single-curve NLS calibration with multi-start Levenberg-Marquardt and AIC + eligibility-gate selection.
