Limitations and responsible use

• The implementation targets static, baseline-covariate prediction.
• It returns one RMST prediction per chosen horizon, not a coherent full curve.
• The marginal Kaplan-Meier pseudo-observation construction requires stronger censoring assumptions than conditional independence given covariates.
• Predictions are prognostic, not causal. They do not estimate treatment effects or justify treatment selection.
• Grouped agreement and discrimination do not establish individual calibration or clinical deployment utility.
• Hosted backbones transfer processed data to third-party infrastructure.
• Foundation-model packages and checkpoints evolve; reproducibility requires explicit version and checkpoint pinning.

Use independent validation, censoring-support diagnostics and a prespecified horizon before any disease-specific application. Prospective validation and institutional review are required before clinical use.