| name | simulation-test-plan |
| description | Design a compact parameter-recovery simulation plan for the freqTLS 4PL thermal-load-sensitivity model. |
Simulation Test Plan
Use this skill when planning the simulation evidence for a freqTLS slice
(recovery, coverage, edge behaviour, or benchmark sanity). Curie leads; Fisher
reviews the inferential targets.
Procedure
- State the estimand or invariant: recovery of
CTmax, z, low, up, k,
phi; profile equivariance ci_z == exp(ci_log_z); |D(MLE)| ~ 0; or a
warning on a weakly identified design.
- Construct data from known settings with
simulate_tls() (locked
data-generating process, fixed seed, factorial temperature x duration grid).
- Fit with the public
fit_tls() workflow and form intervals with
confint(method = "profile").
- Check estimates and intervals on the scale the user interprets (degrees for
CTmax, degrees-per-decade for z, probabilities for low/up).
- Add at least one scientifically likely edge case (sparse design, no/all
mortality, threshold never crossed,
phi near the binomial limit).
- Keep routine tests deterministic and small; put long studies in
data-raw/.
freqTLS Targets
- Recovery tolerances (guide, from the SPEC):
CTmax to about 0.4 deg C, z
to about 0.6, low/up to about 0.05, k to about 30% relative; wider for
beta-binomial.
logLik(beta_binomial) > logLik(binomial) and AIC(beta_binomial) < AIC(binomial) on overdispersed data; near-binomial on clean data.- Benchmark sanity: cached bayesTLS vs live freqTLS within a loose tolerance
(
CTmax ~ 1 deg C, z ~ 25%); no Stan in the test.
Edge Cases
Too few temperatures or durations; no mortality; all mortality; threshold never
crossed; asymptote not approached; CTmax extrapolated; non-closing profile
(warning + NA endpoint, no crash); grouped designs with shared shape.
