Floating-Point Policy

Floating-point width is selected by ownership boundary and numerical role. A single global dtype would either double host bandwidth or weaken Firth solver stability.

Contract

1. Phenotypes, covariates, LOCO predictions, dosages, and genotype summary buffers are stored as f32 on the Rust host.
2. JAX score-test arithmetic and public result statistics use float32.
3. Null Firth, approximate Firth, fallback Newton-Raphson, likelihood, information-matrix, and corrected-statistic calculations use JAX float64.
4. Corrected Firth values are narrowed once when merged into float32 score results.
5. firth_dtype does not exist. Firth width is an algorithm invariant, not a user option.
6. Epsilon is derived from the active array dtype with jnp.finfo(dtype).eps. A float64 epsilon must never be cast down and used as a float32 tolerance.

Configuration Values

Rust configuration thresholds are finite validated f64 newtypes:

• positive tolerances and variance floors are greater than zero;
• probabilities are in (0, 1) and probability floors in (0, 0.5);
• step-halving scales are in (0, 1);
• sparse carrier dosage thresholds are in (0, 2].

The PyO3 boundary extracts the validated scalar with .get() and passes a Python float into the typed JAX configuration. Kernels cast the scalar to the array dtype at the point of use.

Review Checklist

• New host arrays remain f32 unless a measured host-side algorithm requires wider storage.
• New Firth operands are explicitly converted to jnp.float64 before solver work begins.
• New convergence constants derive from the operand dtype.
• No configuration or manifest field reintroduces firth_dtype.
• Narrowing output conversions occur once during materialization, not inside iterative kernels.