benchmark-vs-bayestls-audit

name	benchmark-vs-bayesTLS-audit
description	Audit the freqTLS-versus-bayesTLS benchmark for a fair configuration, sound cache provenance, and the correct R-SHRIMP data rebuild.

Benchmark vs bayesTLS Audit

Use this skill for any change to data-raw/build_benchmark_cache.R, the vendored datasets, R/data.R, or the comparing-to-bayesTLS vignette. Jason maps the comparator; Rose audits the claims and provenance.

Fair-Configuration Gate (R-RELABS)

The two interval-bearing model fits must use the relative midpoint threshold and the constant-shape configuration so their comparison is fair:

• classical two-stage: bayesTLS::ts_stage1 -> ts_stage2 -> ts_ci (absolute LT50 by construction; label it as an approximate comparator and use it only where fitted lethal asymptotes near zero and one make the thresholds close);
• Bayesian: bayesTLS::fit_4pl(temp_effects = "mid") -> extract_tdt(target_surv = "relative");
• freqTLS: fit_tls() -> confint(method = "profile").

Confirm the model temperature effect runs through the midpoint only (shared low, up, k), the time unit and tref match across all three (pin t_ref/time_multiplier when calling bayesTLS; R-UNITS), and the comparison reports point estimates plus CI width and asymmetry. Never describe the classical estimate as using the relative threshold. Zebrafish is a 3x3 per life stage. Include the fairness footnote.

The R-SHRIMP Data Fix

The bayesTLS shipped shrimp_lethal death counts are corrupted upstream: the CSV column Mortality_after_trial is a proportion (e.g. 0.0909 = 1/11, 0.5 = 5/10), and the upstream make_datasets.R mislabels it a count and applies as.integer(...), flooring proportions below 1 to 0. freqTLS must rebuild the counts from the CSV:

• deaths = round(prop * N), hence survived = N - round(prop * N);
• assert the rebuilt death distribution is sane (not collapsed to all zero);
• document the rebuild in R/data.R and a friendly upstream report;
• verify the rebuilt .rda before finalising.

Zebrafish counts ship correctly (n_surv, n_total); rename to native columns only.

Cache Provenance Gate (R-STALE)

Stan will not run in CI, so the benchmark reads a maintainer-built cache at inst/extdata/bayesTLS_benchmark_cache.rds. Confirm the cache meta records bayesTLS_version, a verified 40-character git_sha, source_url, cmdstan_version, date_built, seed, the configuration, and the R-SHRIMP note. The vignette shows the live calls with eval = FALSE, reads the cache with eval = TRUE, runs freqTLS live, and prints the provenance. A test-benchmark-sanity tripwire checks the cached numbers against a live freqTLS fit within a loose tolerance and a one-command regeneration path exists.

Licence Gate (R-LICENSE)

Licensing is component-specific. Confirm each installed file appears in the data licence ledger with source-specific terms, attribution in R/data.R (@source), inst/CITATION (freqTLS + bayesTLS bibentry), and the README where relevant, with code under GPL-3 and provenance in inst/COPYRIGHTS. Permission-pending snow-gum and Kristineberg material must remain build-excluded.