The Signature You Cannot Repeat

The check

Running independent failure cases.

Every uncertainty and free choice

• Sampling: Pointer Events plus getCoalescedEvents() when present. Browser, device, operating system, display load and input type change the samples.
• Time: each event batch is mapped onto the current performance.now() clock while retaining coalesced-event deltas. Duplicate times remain; every derivative uses a displayed 1 ms floor.
• Space: CSS pixels, not tablet units. Translation can use the path-length-weighted centroid, first point or nothing. Isotropic scale uses the bounding-box diagonal. Rotation is optional and off.
• Resampling: each stroke is resampled separately at equal arc length, with 128 points total by default and at least two per stroke. The two-per-stroke floor wins, so a trace with more strokes than half the requested count ends above that count; the transform ledger prints the requested and allocated totals for both traces. Pen-up travel is excluded.
• Size limits, and one rule changed after seeing data: a file may not exceed 2,000,000 bytes or 20,000 samples per trace. A review specimen inside both limits, 10,000 two-point strokes per trace, would have been resampled to 20,000 points and asked for two 400,000,000-cell matrices, about 6.4 GB. The parser now also refuses more than 512 strokes in a trace, which holds the warping matrix at or under 1024 by 1024 cells. That limit was added after seeing that specimen.
• Smoothing: optional five-sample Savitzky-Golay on x and y with coefficients [-3, 12, 17, 12, -3] / 35. The first and last two points of each stroke are unchanged.
• Derivatives: direction is [dx/r, dy/r]. Speed is r / max(dt, 1 ms). Acceleration is the speed difference over the same floored time and is too jitter-sensitive for the default score.
• Standardization: speed and pressure use one pooled median and median absolute deviation for the pair, with epsilon 0.000001. Pressure defaults to weight zero. The page forces that weight to zero unless both traces contain genuinely supported pen-pressure samples.
• When the epsilon becomes the divisor: if both marks are drawn at an almost perfectly steady rate, the pooled speed spread falls below that epsilon and the standardized speed term is divided by the epsilon rather than by the data. The score can then reach the millions while the two performances are in fact very close. The page detects this case and prints the caution beside the score instead of letting a large number speak for itself.
• Local cost: wDir*((sinA-sinB)^2+(cosA-cosB)^2) + wSpeed*(vA-vB)^2 + wLift*(bA-bB)^2 + wPressure*(pA-pB)^2.
• Warping: C[i,j] = d(i,j) + min(C[i-1,j], C[i,j-1], C[i-1,j-1]), starting at the top left and ending at the bottom right. The default has no Sakoe-Chiba window. The window is not decoration: on the shipped pair the score moves from 0.421029 unwindowed to 0.430038 at 8, 0.476719 at 4 and 0.633974 at 1. It cannot change the counterfactual, because both sides of that comparison are resampled at equal arc length and end up the same length, so the diagonal alignment is always available at zero cost. A comparison run on raw captured samples rather than arc-length resampled ones does lose its timing blindness as the band narrows, which is one more reason the resampling step is named here rather than buried.
• Stroke constraint: different stroke indices have infinite local cost unless the reader disables the constraint. Different stroke counts can therefore produce no finite path, which the page states instead of printing a score.
• Reported score: accumulated cost divided by recovered path length. It is a weighted-feature score: every nonzero weight, including the stroke-start weight of 4, is inside the headline number, which is why the headline names the weights instead of naming two features. Settings where no data-varying feature carries weight are forbidden.
• A second rule changed after seeing data: the weight guard above asks whether a feature reads the drawing at all. A reconcile pass measured that this is the wrong question for the counterfactual column, because the uniform-time probe copies direction, stroke start and pressure from its source, leaving speed as the only channel that can differ from it. Every one of the sixty reachable weight settings with a zero speed weight therefore scored exactly zero against that probe for every mark tested, while the guard reported no adjustment. The page now measures the condition directly, by asking whether every channel that carries weight is pointwise identical across the pair, and says so beside the score instead of letting a forced zero read as agreement. The guard is not widened to change the reader's weights here, because against captured repetition B a zero speed weight is a perfectly informative setting; only against the copied probe is it empty.
• Privacy audit: the counter covers page-level calls to fetch, XHR open and send, WebSocket, EventSource, beacon, form submission and service-worker registration, plus Resource Timing entries after load. It cannot audit extensions, the browser, the operating system or the host.