The simulation holds the 1,200 supplied covariate rows fixed and replaces every outcome in every replicate. This resolves the unspecified covariate generator in the initial specification without using fitted outcome associations.

Age must be from 18 through 40 inclusive. Hot days must be from 0 through 20 inclusive; fractional values are permitted. Smoking must be 0 or 1, and birthweight must be positive. All required fields must be finite and complete, and ids must be unique. Saved boundary checks exercise both inclusive endpoints, values just outside each range, and a valid fractional exposure.

All contrasts standardize over full rows. Age-specific rows retain observed smoking values and recalculate age_c after replacing age. The squared age term is evaluated inside the model formula. Contrasts use marginaleffects with the supplied sandwich HC3 matrix, central numerical differentiation, and df = Inf for normal intervals.

The generator has mean mu and conditional variance mu^2 * (exp(0.01) - 1). Its positive support is structural when mu is positive. One million independent multipliers check the intended moments; one dataset per world checks support, dispersion, and range before repeated fitting. Dispersion ratios outside 0.5 to 2 or more than 5% outside the supplied outcome range produce warnings, not a change in the declared world.

Monte Carlo rate intervals use the two-sided 95% Wilson binomial interval. Its width remains positive when an observed rate is zero or one. Bias and RMSE uncertainty use Monte Carlo standard errors. Metrics use successful runs; counts and bounds over all attempted runs retain failures in the accounting.

RNG seeds are recorded per replicate. Moment validation uses 20261130. Generator examples use 20262001 through 20262003. Repeated outcomes use 20270000 + 10000 * scenario_index + replicate, with scenario_index 1 = null, 2 = main, 3 = curved. R uses its default Mersenne-Twister generator and Inversion normal generator.

The lockfile was generated by renv static dependency discovery from the installed packages used in execution. It contains 49 packages and the complete recorded dependency closure. renv status reports a consistent state against that installed library. A clean restore into a new isolated library was not run.

Reproduction commands are just check, just simulation 400, just analysis, and just figures. Run them from the project root. All R commands pass through bin/rscript-pty. The simulation runs independently of the production fit and contrast scripts. The wholly synthetic exercise makes no claim about real pregnancies.
