From a reading to a research record
What the checks
can tell us.
A score can play correctly as a file and still contain a misread note. This workspace makes the different kinds of evidence visible.
- 01
Keep the source in view.
The printed page is the starting point. A visual reading identifies notes, rhythms, clefs, ties and annotations. The surrounding dialogue helps explain what the example is doing.
The reading edition records VLM-assisted transcription. The species corpus began with manual encoding. Those histories are kept separate.
- 02
Make the reading editable.
The corpus stores canonical
**kernscores with JSON metadata. The reading edition uses MusicXML and source-linked records. A rendered score makes an encoding easier to inspect.A concrete example: the three-voice pilot
build_content.pygenerates MusicXML from a structured score ledger. It also compares printed interval numbers with the diatonic intervals implied by the notes. That comparison can reveal a mistake in the reading or a disagreement in the source itself. - 03
Ask precise, limited questions.
Does conversion preserve the written events? Does a tie have an ending? Do the parts occupy the same amount of time? Can the player read the file?
Format comparisonThe corpus checker compares pitch, onset, duration and rests. The stricter profile also checks ties and pitch spelling. It does not compare voice assignment or engraving.Structural checksThe Fux diagnostic script flags near-matches to recurring cantus firmi, unequal part lengths, unmatched ties and invisible gaps. A flag is a reason to inspect, not an instruction to rewrite.Parsing and renderingThis workspace checks current file bytes, runs the existing playback parser and generates a preview. Saved schema results are shown separately; they are not a new schema-validation run. - 04
Return to the print.
A disagreement needs interpretation. It may be a transcription error, a conversion difference, a source error or a passage that remains unclear. Deliberately faulty counterpoint should remain faulty.
See a clef that remains under review →
What this snapshot contains
The corpus comparison covers 123 paired files. With ties and spelling checked, 113 pairs agree and 10 differ. These figures describe agreement between encodings, not recognition accuracy.
The wider reading edition has 538 distinct primary score files. It includes short illustrations, repeated references and long passages. All 123 corpus exercises are now linked using source-page and per-part musical-event evidence. Same-page location alone is insufficient. The closing whole-note/breve differences are retained explicitly; rests, spelling and engraving are outside this identity test.
Original editions’ review labels are displayed as recorded review. They are historical records, not a new independent scholarly certification.
Leave a trace of your review.
Save a question, discrepancy or source-comparison note beside an example. Each note is tied to the score and source fingerprint in this snapshot. If those inputs change, the earlier note is shown as belonging to a different version.
Notes are saved only in this browser. Export them to keep or share them. Importing a conflicting note preserves the one already stored. A note does not change the source score, the edition’s playback status or a publication decision.
You can also export a review task with input locations, hashes, known issues and a suggested result structure. This prepares a source-comparison job; it does not claim a model has run.
Reproduce and extend the workspace
Import the pinned species checkout, rebuild the catalogue, then run the browser and data checks. The scripts and commands are documented in research/fux/README.md in the Musicbook repository.