The transcription PDFs are vector drawings, so nothing is guessed from pixels. Each cell is a shape with an exact colour, and every digit, letter and label is real text with a position. The reader finds the rows of 16 cells, reads the marks in and above each cell, groups rows into bars and parts, then writes an EVA project and checks it against the app's format.
Your PDFs never leave this browser.
| On the sheet | Read as | Step field |
|---|---|---|
| black / grey cell | hit, level | velocity 110 / 65 |
AC row | accent | accented on that step |
| digit 2–8 | roll | substeps |
| gradient cell + digit | roll getting softer | fading substep velocities |
F | flam | velocity ½, retrig 1/32 × 2, fade +100, nudge −12 |
1/32 + L1.5 | retrig | retrig, retrigCount |
+1/48 over a shifted cell | micro-timing | nudge in 1/384 notes (+8) |
56% over a half cell | gate length | decay 56, synth follows the gate |
| circled slash ⊘ | mute | cut |
| ↑ / ↓ | pitched | pitch ± 2 |
| zigzag | longer decay | decay 200 |
R | reversed | reverse (sample rows) |
| closed + open hat together (808) | third hi-hat | Mid Hat track with a combo rule |
Identical bars become one bar played N times. Bars that differ only by cells in one bar become one bar with conditions, so a ghost note in the last of two bars becomes 2:2.
Each sentence is matched against fixed rules: kit, choke, no-choke, sample length, high-pass, reverb, delay, auto-pan, combine and layer. Anything else is listed as not applied, never guessed.
Screenshots and scans (no vector shapes), sections that aren't "Part N", and prose about a single step. The report lists each one.