Find the recurring differences.
Compare early and late strokes across local drawing-machine records and your imported records, then inspect the patterns behind each result.
Your corpus
Select the records to analyze. Remove an example to hide it, or restore the examples later. Removing an import clears it from this tab; the original file stays on your device.
| Drawing | Strokes / phase | Mean length: early → late | Centroid: early → late | Mark-ID difference | Inspect |
|---|
Each result compares the first and last equally sized groups of accepted strokes; the middle is excluded. Length is the absolute difference in mean path length in canvas units; start-centroid distance uses coordinates normalized by canvas width and height. Mark-ID difference is total variation distance. The table ranks by the chosen effect size, not p-value. Early and late centroid coordinates are shown as fractions of canvas width and height.
Each raw p-value uses a seeded permutation test with the plus-one correction. The displayed q-values are Benjamini–Hochberg adjusted across all tests in the current batch (three metrics per record). The records may be related variants, and the tests are exploratory: adjusted p-values do not establish independent replications or causality. Comparisons of lengths across different canvas resolutions are in source canvas units, not physically calibrated units.
The bundled records are generated drawings from the drawing-machine project. Example selection is stored in this browser when available. Imports stay in this tab and are not sent anywhere; add them again after closing it. Removing a record does not delete its source file. An uploaded record without canvas dimensions uses its largest recorded x and y, so provide W and H for valid cross-record normalized spatial comparisons. The JSON export includes source paths, settings, phase indices, observed values, p-values, q-values, and shuffle distributions.