Changelog

Source: NEWS.md

freqTLS 0.1.0 (experimental release candidate)

CRAN release: 2026-07-21

• Rebased the empirical teaching plan on the pinned bayesTLS supplement rendered 2026-07-14 (commit 76510412e06c594c96894a1baba1f0e1a34a5aea). Canonical cases are oxygen-gradient zebrafish, cereal aphids, Snow-gum PSII, and the mortality and awake/coma Drosophila suzukii endpoints.
• Brown shrimp and life-stage zebrafish remain installed benchmark-only legacy fixtures but are removed from active examples, navigation, and current comparison surfaces.
• Retained Beta responses, limited random intercepts, shape formulas, deterministic heat-injury prediction, the formula interface, and the frequentist interval/diagnostic displays as explicitly experimental.
• Censored-time, hurdle-productivity, fitted repair dynamics, and CRAN submission remain out of scope.
• Corrected formula starts so intercept models initialise only their intercept while no-intercept cell-mean models initialise every cell. Difficult fits may receive a deterministic nloptr Newton refinement when it improves both the objective and maximum gradient.
• Added row-specific parameter prediction for interacted formula designs via predict(..., type = "parameters"), including freq_tls S3 dispatch.
• Added the reviewed canonical bayesTLS comparator cache built on Totoro from pinned bayesTLS commit 76510412, with exact analysis hashes, sampler diagnostics, source versions, seeds, and a public table of actual freqTLS-minus-bayesTLS differences. The historical shrimp/life-stage cache remains internal legacy evidence only.
• standardize_data() now warns with the affected count and epsilon whenever a continuous proportion is moved off 0 or 1 for the Beta likelihood. The Snow-gum example makes its 90 adjusted values explicit.
• Omitted tref / t_ref now resolves to one physical hour for standardized data with recognised duration-unit metadata (for example, 60 minutes or 1 hour). Bare data retain the historical one-native-unit fallback with a warning, so an ambiguous CTmax reference is never silent.
• get_ctmax(), get_z(), and get_shape() now pass method = "wald" or "profile" through to tidy_parameters() and retain per-coordinate interval routing in interval_type.

freqTLS 0.1.0 (unreleased historical candidate)

CRAN release: 2026-07-21

• Corrected grouped contrast direction so dCTmax:A-B, dlog_z:A-B, and dz:A-B now follow their written meaning: group A minus group B.

freqTLS is the frequentist counterpart to the Bayesian bayesTLS package: it fits the four-parameter logistic thermal-load-sensitivity (thermal death-time) model by maximum likelihood via TMB, parameterised directly in CTmax and thermal sensitivity (z). Under the matched relative-threshold, constant-shape configuration, it targets the same fitted curve as bayesTLS; uncertainty is reported through a frequentist trio — Wald (delta), profile-likelihood, and bootstrap — instead of a posterior. Forked from profileTLS (commit 6f963a9, v0.3.3), which it supersedes.

bayesTLS-analogue API (historical 0.1 surface)

• standardize_data() — the shared raw-data entry point for count or continuous-proportion responses (adopted from bayesTLS).
• fit_4pl() + make_4pl_formula() — the direct CTmax/z formula interface (ctmax/z/up/low/k/by, plus threshold, t_ref, bounds, family), fitted by maximum likelihood through the TMB engine; returns a freq_tls workflow object.
• tls() / tls_z() / tls_ctmax() / tls_tcrit() — z, CTmax, and T_crit with confidence intervals at the relative midpoint, the absolute (LT50) threshold, or any LTx.
• extract_tdt() with get_z_* / get_ctmax_* / get_tcrit_* accessors — the nested z / CTmax / T_crit structure, with parametric-bootstrap replicates as the frequentist analogue of posterior draws.
• predict_survival_curves() — the fitted survival surface with bootstrap bands.
• diagnose_tdt_fit() and tdt_parameter_table() — convergence diagnostics (optimiser/Hessian/gradient) and the 4PL parameter table.
• two_stage (ts_stage1() / ts_stage2() / ts_ci() / ts_curve()) — the classical two-stage comparator, reporting both normal and small-sample t intervals.
• The plots (plot_confidence_eye(), plot_survival_curves(), plot_tdt_curve(), plot_heat_injury()) and extractors accept the freq_tls workflow object.
• Shared empirical data include aphid_tdt and zebrafish_o2; the active teaching set and legacy boundary are listed above.

Inference and calibration

• Small-sample Bates–Watts profile-t / Wald-t calibration: confidence intervals reference a t distribution with residual df = n − p, restoring nominal coverage at small n and reducing to the asymptotic interval as n grows. The evidence is a coverage + width simulation (repository-only data-raw/calibration-study.R, not installed): at df ≈ 10 the asymptotic 95% interval covers ~0.93 and the t-correction restores ~0.96.
• A three-way benchmark (repository-only data-raw/benchmark-vs-bayes.R, not installed): freqTLS reproduces bayesTLS’s CTmax to ~0.07 °C on the brown-shrimp data, beside the classical two-stage estimator.

Analogue S3 surface

• confint(), summary(), ranef(), and coef()/logLik()/vcov()/nobs(), the heat-injury functions (predict_heat_injury() / plot_heat_injury() / heat_injury_envelope()), and check_tls() all accept the freq_tls workflow object — this listed post-fit surface works on the fit_4pl() result.
• fit_4pl(by = "g") now labels groups by the bare factor levels (CTmax:young_embryos), identical to the column interface, end to end.

Case studies and vignettes (render without Stan)

• The frequentist-and-bayesian centerpiece carries the coverage panel and the three-way benchmark; comparing-to-bayesTLS carries the live + cached comparison.
• Historical articles included unpublished compatibility fixtures. They are not part of the active teaching set.

Simulation

• The build-excluded repository directory scripts/simulations/ contains a freqTLS (ML/TMB) analogue of the bayesTLS two-stage-bias simulation (shared data-generating process + scoring), with a comparison to the bayesTLS results. These maintainer scripts and their DRAC launcher are not installed with the package.

Engine

• Asymptotes use bayesTLS’s disjoint-bounds reparameterisation (compute_4pl_bounds); up is now a direct coordinate.
• Convergence / positive-definite-Hessian status is surfaced at fit time.

Superseding profileTLS

• freqTLS is the intended successor to profileTLS, its engine donor.