unified_symbolic_rep — demo pages

Time-aligned self-supervised representation learning for solo-piano MIDI (scores and performances). Each demo is one page under this root; pages are added as stages complete. The MIDI players need internet (CDN) and a hosted soundfont.

Stage 1 — local encoder

• Reconstruction demo (s1a_v0 step 26000; 6 val pieces from POP909 / Aria-MIDI / ATEPP, original vs reconstruction, 2026-09-02)

Stage 2 — contextual trunk

• Contextual encoder mechanism (figure: frozen local encoder, masked-latent trunk, visible-only student and EMA teacher)
• Flattened performance view (J5c) (a performance and its deterministic score-like twin: chord snap to frame, one velocity, pedal/legato undone, rubato kept; 2 ASAP performances beside their DCML score render + 4 val performances, 2026-09-12)
• Humanised score view (J5c view B for scores) (rule-based re-rendering of score renders at ×1.0 and ×1.5 deviation scale beside the real performance for 2 Beethoven sonatas + 4 val scores; no tempo curve, 2026-09-12)