Skip to content

Validation report

Regenerate this page with python tools/validation_report.py. CI runs the same script on every push and fails the build if any feature drifts outside tolerance or disagrees with the reference on NaN, so a stale or optimistic table here cannot survive a merge. The run below is from a maintainer machine; the CI run publishes the same table as a job summary and an artifact.

kymora 0.2.0 on Windows-11-10.0.26200-SP0, Python 3.14.6, numpy 2.4.6.

Each of the 33 features is compared against a numpy/scipy reference implementation across 7 sample series (gaussian, trend, sine, constant, tiny, single, integers_with_ties), covering normal, trending, periodic, constant, two-element, single-element, and heavily-tied data.

Tolerance: atol=1e-10, rtol=1e-09. NaN must agree exactly -- the NaN policy is part of the API contract, so a NaN where a number was expected (or vice versa) fails regardless of tolerance.

feature max abs error worst-case series series compared verdict
mean 2.842e-14 trend 7/7 within tolerance
std 7.105e-15 trend 7/7 within tolerance
var 4.547e-13 trend 7/7 within tolerance
min 0.000e+00 - 7/7 bit-identical
max 0.000e+00 - 7/7 bit-identical
median 0.000e+00 - 7/7 bit-identical
quantile_10 0.000e+00 - 7/7 bit-identical
quantile_25 0.000e+00 - 7/7 bit-identical
quantile_75 0.000e+00 - 7/7 bit-identical
quantile_90 0.000e+00 - 7/7 bit-identical
skewness 2.671e-15 trend 5/7 within tolerance
kurtosis 8.882e-16 trend 5/7 within tolerance
abs_energy 4.657e-10 trend 7/7 within tolerance
root_mean_square 1.421e-14 trend 7/7 within tolerance
mean_abs_change 4.441e-16 trend 6/7 within tolerance
mean_change 0.000e+00 - 6/7 bit-identical
cid_ce 7.105e-15 gaussian 6/7 within tolerance
mean_second_derivative_central 8.890e-18 gaussian 5/7 within tolerance
zero_crossings 0.000e+00 - 7/7 bit-identical
mean_crossings 0.000e+00 - 7/7 bit-identical
number_of_peaks 0.000e+00 - 7/7 bit-identical
longest_strike_above_mean 0.000e+00 - 7/7 bit-identical
longest_strike_below_mean 0.000e+00 - 7/7 bit-identical
autocorr_lag_1 8.882e-16 trend 5/7 within tolerance
autocorr_lag_2 5.551e-16 trend 4/7 within tolerance
autocorr_lag_5 2.220e-16 trend 4/7 within tolerance
autocorr_lag_10 5.551e-16 sine 4/7 within tolerance
trend_slope 4.337e-19 integers_with_ties 6/7 within tolerance
trend_r2 2.220e-16 trend 5/7 within tolerance
permutation_entropy 1.110e-16 integers_with_ties 5/7 within tolerance
dominant_frequency 0.000e+00 - 5/7 bit-identical
spectral_centroid 3.886e-16 gaussian 5/7 within tolerance
spectral_entropy 6.661e-16 gaussian 5/7 within tolerance

Rows reading NaN by contract are features that are undefined for every sample series in this matrix (none currently) -- they are still asserted to be NaN in both implementations.

All features within tolerance.

How to read this

max abs error is the largest absolute difference between the Rust core and the numpy/scipy reference over every sample series where both produced a finite number. series compared is how many of the sample series contributed — a feature that is NaN by contract on constant or very short series compares on fewer of them, and its NaN agreement is checked separately and exactly.

bit-identical means the two implementations produced the same float bit for bit. Everything else is float-summation ordering: the Rust core sums in a different order than numpy, which is why var on a trending series differs around 1e-13 while min never differs at all.

The tolerance is the same one numpy.testing.assert_allclose applies in the test suite, atol + rtol * |expected|, so this report and pytest tests/ agree by construction rather than by coincidence.

What this does not cover

This is a numeric accuracy report over a fixed sample matrix. Robustness — that no input crashes the extension or violates the NaN policy — is a separate concern covered by the hypothesis property tests in tests/test_properties.py, and the NaN contract itself is pinned by tests/test_nan_policy.py. Neither substitutes for the other.