Every test digit the hundred-template classifier gets wrong, with the fit animated. Two curves deform at once: the best template of the digit's true class, and the best template of the class that actually won. Watch them for a moment and something looks off — the blue curve, the one belonging to the right answer, usually settles onto the digit better. That is not an illusion, and the numbers under each card say why.
In the staged fit the knots, warp, truncation and pen width are held still for the first half of the run, so only translation, rotation, scale and shear move; then everything is released. Staging changes the path rather than the destination, and since the score is a time average along that path, it changes the score — by 0.925 to 0.905, slightly for the worse. These are the digits where that shows.
Built by python animate.py then python build_page7.py. The classifier is the one described in classify10.html.