[vkosuri]

Adds "Seven ideas, in motion" after "How it works" on the playground: seven
looping SVG sketches for the terms new readers ask about.

• Offline vs online handwriting: a scan, a tablet and the sensor pen
• Writer-independent and writer-dependent (72.46% vs 80.07%)
• Uncertainty: five runs agree, disagree, or all hesitate between o and O
• Left-handed writers: pull vs push, 74.92% vs 59.47%
• Small or capital: 43% of errors, 84.3% case-insensitive
• Why it gets it wrong: missing position (sensor), examples (data) or method (algorithm)
• What raised accuracy (69.2% to 74.46%) and what did not (bigger model, 70.4%)

A second section, "How CNN, LSTM and hybrid neural networks read handwriting",
has five more: a neural network reading and learning, a CNN filter, an LSTM
and a BiLSTM, attention, and the CNN-BiLSTM-attention hybrid. Model sizes
match imu2text/models.py. For search: question headings, a nav link, CNN
and LSTM in the meta description, about/teaches/keywords in the
JSON-LD, and a sitemap lastmod.

CSS only, --v-* tokens only, no inline styles (the served CSP drops
them). Scene buttons stop on one scene; with prefers-reduced-motion
nothing loops.

Please check:

• every number against docs/benchmarks.md and docs/uncertainty.md, and the
  model sizes against imu2text/models.py; the drawings are sketches and the
  intros say so
• the "left hand pushes" framing reads fairly
• the section on a phone and with reduced motion turned on

🤖 Generated with Claude Code