Use of generative AI

comprisk is developed with Claude Code, Anthropic's command-line coding agent, using Claude Opus and Claude Sonnet models. It works from written specifications, edits files in this repository, and runs the test suite and the benchmarks itself. The authors read its diffs and decide what merges.

Roughly 80% of the test suite, 50% of the library source and 90% of the documentation began as agent drafts, and agent-drafted code appears in every module. The authors decided the statistical methodology, the choice and formulation of each estimator, and the validation described below.

Two mechanical checks run on top of that review. The regression and non-parametric estimators are gated on R-generated fixtures committed to the test suite, and the forest is validated against randomForestSRC through a maintainer-run harness under validation/alignment/. A regression in any of them fails CI against numbers cmprsk, crrp, survival, riskRegression and randomForestSRC produced.

The authors understand this implementation and take responsibility for the correctness of the software and the accuracy of the claims made about it.