research_tools

Fifty-eight helpers for numeric research data. No scipy, no pandas, no install.

Design rule. Every function is short enough to read in full before you
trust it, returns a plain dict or array, and never touches the network or the filesystem. That is the
whole reason for not using a heavyweight library: when a number matters, you should be able to audit
the ten lines that produced it.

import research_tools as rt
v = rt.parse_numbers("4.12, 4.35, 4.40, 4.55")
rt.describe(v)        # -> n, median, MAD, range
rt.grubbs(v)          # -> G, critical value, is-outlier
rt.bootstrap_ci(v)    # -> 95% interval, 2000 resamples, fixed seed

Robust statistics

`median`	`v`
`mad`	`v`
`iqr`	`v`
`robust_z`	`v`
`winsorize`	`v, p=…`
`trimmed_mean`	`v, p=…`
`skew`	`v`
`kurtosis`	`v`
`cv`	`v`
`describe`	`v`

Outliers and randomness

`outliers_mad`	`v, k=…`
`outliers_iqr`	`v, k=…`
`grubbs`	`v, alpha=…`
`runs_test`	`v`

Distributions and tails

`fit_lognormal`	`v`
`hill`	`v, k=…`
`pwm_tail`	`v`
`ks_uniform`	`v`
`empirical_cdf`	`v, x`
`entropy_hist`	`v, bins=…`

Uncertainty and detection

`propagate_ratio`	`y, sy, x, sx`
`combine_uncertainties`
`welch_dof`	`s1, n1, s2, n2`
`coverage_95`	`dof`
`uncertainty_interval`	`x, u, dof=…, k95=…`
`snr_db`	`signal, noise`
`detectability`	`blank_std, k=…`
`currie_limit`	`blank_mean, blank_std, k=…`

Signal handling

`detrend_linear`	`v`
`moving_average`	`v, n=…`
`median_filter`	`v, n=…`
`autocorrelation`	`v, lag=…`
`hurst_rs`	`v`
`dominant_frequency`	`v, dt=…`
`psd_band`	`v, dt=…, lo=…, hi=…`

Sampling

`gy_mass`	`d_mm, rel_error=…, const=…`
`bootstrap_ci`	`v, stat=…, n=…, alpha=…, seed=…`
`jackknife_se`	`v, stat=…`
`permutation_p`	`a, b, seed=…, n=…`
`sample_size_mean`	`sd, margin, z=…`
`power_t`	`effect, sd, n, z=…`
`split_sample`	`v, seed=…`

Regression and fit quality

`ols`	`x, y`
`theil_sen`	`x, y`
`aic`	`residuals, k`
`bic`	`residuals, k`
`residual_check`	`x, y`
`variance_inflation`	`xs`

Data hygiene

`parse_numbers`	`text`
`round_sig`	`v, sig=…`
`dedupe_exact`	`v`
`find_duplicates`	`v`
`gaps_in_sequence`	`v`
`constant_columns`	`cols`
`missing_report`	`rows`

Provenance

`hash_dataset`	`data`
`provenance_record`	`source, n, note=…`
`checksum_list`	`v`

What is deliberately absent

No p-value machinery beyond approximations, no plotting, no data-frame operations, no automatic
model selection. Each of those pulls in a library whose defaults then decide the answer — which is
precisely the failure the corrections register documents sixteen times.

If a function here is wrong, you can see it, fix it, and say so. That is a property worth
giving up convenience for.

58 functions · 4 November 2026 · used by the
test kit and browser audit.