Determinize #
A probabilistic language whose sample sites are labelled E or G. Determinizing a program turns its E samples into the means of their distributions and keeps its G samples.
A reviewer reads the statements of the theorems and the definitions they use, not the proofs.
lean/README.md explains
what to review and how the development differs from the paper.
Determinize.Theorems: every theorem, stated in full over the definitions inDeterminize.Specand proved by one term fromDeterminize.Proof, the paper's theorems first and under the paper's names. Start here.Determinize.Spec: what the theorems are about, from syntax, typing, the primitive distributions and determinization to the output and trace laws (Determinize.Spec.Traces.Semantics) and expectations and variances (Determinize.Spec.Expectation).Determinize.Proof: the proofs. The build fails if a theorem uses an axiom other thanpropext,Classical.choiceandQuot.sound, or if its statement relies on anything from these modules other than proofs.Determinize.Frontend: parsing, name resolution and affinity inference (Determinize.Frontend.Infer, the subject ofDeterminize.Theorems.inference_correctness).Determinize.FiniteandDeterminize.Checking: exact exploration of finite-state programs, and the checks that turn an explored model and candidate results into a theorem about the program, checked by Lean's kernel (Determinize.Checking.Statistics).
The default build also checks the coverage examples in Determinize.Proof.Examples. The API
documentation covers every module except the command-line entry point Main, the numerical
evaluator in Determinize/Runtime/, and the tests.