dsh-probstat

LLMs have a persistent weakness around probability and statistics: confabulated z-table values, off-by-a-tick quantiles, mismatched critical values, and muddled compound-event identities. dsh-probstat plugs exactly that gap for DeepSeek Harness with a four-tool, zero-runtime-dependency math kit that returns deterministic, verifiable numbers.

The four tools partition the work cleanly: dist_calc handles pdf, cdf, survival, quantile, and summary statistics for six distributions (normal, binomial, Poisson, exponential, uniform, geometric); z_score converts between standard-normal z-scores and tail or inter-valence probabilities; confidence_interval builds mean intervals (z or Student-t) and Wilson score proportion intervals that stay honest on small samples and extreme proportions; event_probability evaluates union, intersection, conditional, Bayes, complement, and at-least-one identities in one call.

Under the hood the numerics are deliberate rather than approximate: the normal cdf uses the Abramowitz and Stegun 7.1.26 erf fit with absolute error below 1.5e-7, the inverse normal relies on the Acklam algorithm with relative error under 1.15e-9, the Student-t is computed via regularized incomplete Beta and Lanczos log-gamma cross-checked against published tables, and binomial and Poisson pmfs are built by exact recurrence to sidestep combinatorial overflow. Every invalid input returns valid: false with a human-readable explanation—nothing ever throws.

This is the tool to wire into a dsh agent stack whenever your workflow demands that a p-value, a credible interval, or a posterior odds is computed correctly the first time: hypothesis testing, Bayesian inference, reliability modeling, or any pipeline where a wrong number is worse than no number at all.