Typing and performance evidence

Use the repository checks to assess the current checkout. Historical source counts, timing measurements and pass counts are not current qualification.

make typing

The typing checker compares diagnostics and records the relevant environment against typing-baseline.json. Investigate drift before changing the baseline. Do not regenerate it solely to silence a failure or weaken types to make a check pass.

The counterfactual benchmark can measure declared workloads. Run each case in a fresh process, compare shapes and numerical outputs before timing, and record versions, cache state, repetitions and competing workload.

Counterfactual batching limits per-call work, not total resident memory. Scenario inputs, posterior draws and complete outputs may remain materialised; native allocations and graph intermediates also contribute to memory use. A synthetic benchmark does not establish a production capacity guarantee or statistical validity.