Nine Apache-2.0 libraries. This repository is the index. It does not compute.
One compares two aligned records. Eight score a lunar site. Each prints a decision and is allowed to refuse it. You bring the measurements. Nothing is downloaded and nothing is hosted. There is no second license for these files.
splitrecord is a trend library for two aligned columns, one number per line. The columns are whatever records you already have. A water level is one use. The code does not know the unit.
Each column becomes a sample z-score: subtract the mean, divide by the sample standard deviation. The divisor is n−1. The residual is z(A) minus z(B). A column with no spread is refused.
The library then prints the size of the drift, a test of whether the residual is monotonic, and 95% limits on that slope. The limits are not the size.
Theil–Sen slope. The size. For every pair of rows i < j, the slope is (vⱼ − vᵢ) / (j − i). The result is the median of those slopes. An odd count takes the middle slope. An even count takes the average of the two middle slopes. The unit is z per row, because the denominator is the row gap, not a date. The estimator does not assume a normal distribution. There is no intercept. The median is a point, not an interval. sen95, hamed95, and gilbert95 are separate ranks of the same slopes. They are not the median. --seasons does not print them. A median is why one wild point does not set the answer. The known breakdown point of this estimator is 1 − 1/√2, about 29%. This library computes the median. It does not run a separate breakdown trial.
Mann–Kendall test. The order. S counts pairs where the later row is higher, minus pairs where the later row is lower. A tie adds nothing. S is a count. It is not a slope and not a probability. For 8 or more rows, a two-sided p-value is the normal approximation with the continuity correction (S − sign(S)) / √var. The variance starts at n(n−1)(2n+5)/18. Each tied group of size t > 1 subtracts t(t−1)(2t+5)/18. That variance is multiplied by the Hamed–Rao factor, so a series that repeats itself is not treated as independent. The factor removes the Theil–Sen slope, ranks what remains, and keeps a lag only when its autocorrelation exceeds 1.95996398454/√n. The autocorrelation uses one mean of the whole rank series and the full sum of squares. Under 8 rows the printed p-value is short. The exact small-sample distribution is not computed. A non-positive corrected variance is dependent. The p-value is not a certificate.
The printed Mann-Kendall line also carries Kendall's tau-b, the corrected variance, n_over_nstar, and z. z is (S − sign(S)) / sqrt(var). p is erfc(|z| / sqrt(2)). On a straight residual, n_over_nstar is 1. Yue and Wang's lag-1 factor is not used.
Sen's 95% interval is two ranks in the sorted pairwise slopes, using the tie-corrected variance with no Hamed-Rao factor. hamed95 is the same rank rule after that variance is multiplied by n/n*.
Under 8 rows with no ties, splitrecord labels the interval sen95=exact. That rank is the exact no-tie Sen interval, not the normal approximation. Sen's seven-point series, times 1, 2, 3, 4, 10, 12, 18 and values 9, 15, 19, 20, 45, 55, 78, has median slope 4 and exact limits 3.714285714 and 4.375. At 8 rows and above the label is sen95=normal.
gilbert95 is the interpolated form of that normal approximation: a straight line between the floor rank and the ceiling rank. It is not the rounded sen95=normal pair, and it is not a table copied from Gilbert (1987).
The other eight lines print the inputs next to the word. A fit line prints value, low, and high. A pin line prints void fraction, slope, and offset. A rim line prints on-rim, slope, setback, and width. A pit line prints mouth, void fraction, drop, and slope. A tube line prints width, length, echo, and clutter. A roof line prints span, roof, tensile, crack, ucs, and lithostatic. A site line says whether the hours were supplied or computed. A walk line prints the dose and the number of edges.
The default line is variance=hamed-rao. That correction is for autocorrelation in one series. It does not stop January from being compared with July. --seasons 12 is the seasonal Mann-Kendall test: each month is compared only with the same month in later years. Its variance is the sum of the monthly variances, and the slope is z per year. --covariance uses the Hirsch-Slack covariance instead. Hamed-Rao is not multiplied on top of either seasonal variance. A short last year makes --covariance refuse the table as uneven.
--prewhiten removes lag-1 only after the Theil-Sen slope is taken off, then puts the slope back. The variance on that line is ordinary. It is not Hamed-Rao, and it is not the seasonal test. A flag that asks for two of these corrections is refused.
The worked pre-whitening file is examples/pw_left.csv with examples/pw_right.csv in splitrecord. Nine equally spaced rows. Sen's slope is the median of the 36 pairwise slopes. The line prints that removed slope and the slope of the blended series as two fields.
| Library | Kind | Question | You type | It prints |
|---|---|---|---|---|
| feasfront | Flat-horizon hours, then a limit check | Does the site clear the limits you named | Latitude, longitude, horizon angle, and duration, or hours you already have, plus the slope | ok, or slope, sun, earth, night, or missing. Sun hours count steps with the solar elevation above the horizon. The subsolar longitude walks 360° in 29.530588853 days. Earth hours are the whole duration or zero. Night is duration minus sun hours. No terrain. No libration |
| dosepath | Dose-capped grid walk | Is there a walk inside the clock and the dose cap | Edges x1,y1,x2,y2,t,cost, plus start, goal, clock, and dose cap |
The nodes, or stay. Orthogonal and diagonal steps each cost one tick. A missing edge, a negative cost, or a non-finite cost is not a step. A non-finite dose cap is a stay. Not a radiation model |
| fitslip | Parts envelope | Is every part inside its limits | name,value,lo,hi, or a values file plus name,lo,hi |
pass, fail, or unknown. A blank, a non-finite number, a reversed limit, or a name missing from the envelope is unknown. An empty bill fails. Fail beats unknown. It does not touch the part |
| pinfault | Pin clearance | Can the pin stay | Void fraction, or n_invalid and n_cells, plus slope in degrees and offset in meters |
voids above 0.15 first, even if the slope is negative. Then missing for a fraction outside 0 to 1, a negative slope or offset, a non-finite number, or a missing slope or offset. Then slope above 20°, then offset above 30 m, else ok. Ok is not a clearance to land |
| rimkeep | Rim road gate, plus a sphere distance | Is the rim a road | on_rim, inner slope, shadow setback, width |
on_rim if the flag is true. A present negative or non-finite slope, width, or setback is missing before the setback gate. Then psr if setback is present and under 50 m. Then missing if slope or width is absent. Then slope over 15°, then thin under 30 m, else ok. Overlap is a haversine on a sphere of radius 1,737,400 m. Ok is not a road |
| baghold | Pit mouth gate | Is the mouth usable | Mouth, floor void fraction, optional drop, floor slope | missing, then pinch under 30 m, then voids over 0.15, then drop over 30 m, then slope over 20°, else ok. A missing drop is not a fail. A skylight is not a shelter |
| tubewalk | Tube line gate | Is the line a tunnel | Width, length, echo, clutter, or a point list | dark, then missing, pinch under 10 m, stub under 30 m, clutter, else ok. section segments points within 15 m, fits a circle per segment, and prints the path, the closure against the chord, and the radial rms. las reads LAS 1.2 formats 0 through 3 and LAS 1.4 formats 0 through 10, through the XYZ, return, and class prefix. A waveform packet is not decoded. laz reads the chunk table through laspy and lazrs. Points inside a .laz file are not decoded here. Cloth labels use LAS 1.4 names: class 2 ground, class 7 low point, class 18 high noise. The point-14 class context is (previous class mod 32) * 2, plus 1 for a single return. ok is not a ceiling |
| archhold | Roof thickness, plus a closed-form rock check | Does the roof pass the shape check | Span, thickness, tensile strength, a crack flag, optional depth | missing, then thin under 2 m, then wide over 5,000 m, then weak under 1 MPa, then crack, then load if the lithostatic stress 3100×1.62×depth/10⁶ MPa exceeds the Hoek–Brown unconfined mass strength, else ok. archhold mesh repeats that stress on a grid whose edges do not carry load. archhold fea stores a plastic strain and line-searches an elastic step. The returned stress is in equilibrium up to load factor 0.3608335853. At the full burial load the imbalance stops at 5.911552654. digest is SHA-256 of that line. A person signs it. It is not ABAQUS and it is not UDEC. ok is not a keep |
git clone https://github.com/DigitalCurrensy/splitrecord.git
cd splitrecord
pip install -e .
python -m unittest tests.test_kernel
python -m splitrecord examples/left.csv examples/right.csv
The other eight use the same install and the same unittest command. The sample file is under examples/ in that repo. Python 3.11 or newer. Eight of the libraries have no third-party packages. tubewalk is the exception: pip install -e . also installs laspy and lazrs.
Rows in examples/ were typed to prove the rule. They are not a customer file.
Apache-2.0. Copyright 2026 Digital Currensy Inc. LICENSE is the unmodified Apache text. NOTICE holds the copyright.
These repositories are public. This index does not host a service.