The exemplars

The exemplars are two arrangements of Battle Hymn of the Republic for solo piano. They play side by side on the landing page. Each step that produced them is recorded, and each is checked by something other than the step itself.

The source

The project transcribed the 1862 Ditson edition from its scans. The transcription is research/battle-hymn/battle-hymn-ditson-1862.abc (12,766 bytes). It is the only source the arrangers were allowed to follow. Two readings in the scans are ambiguous; the reference transcription records one choice for each, and the arrangements are independent works that may diverge from it.

The arrangers

Each arrangement is the answer to one call on Ollama Cloud, on 2026-09-26, at temperature 0, with the same system prompt, the same brief and the same transcription. glm-5.3’s first call, with thinking set to high, spent its whole output allowance thinking and returned no score, so its arrangement comes from a second call, at medium. The brief names the transcription as the only source and forbids anything taken from a later arrangement.

	glm-5.3	kimi-k3
Thinking	medium	high
What the project changed	Removed the Markdown code fence around the answer	Added \language "english", because the model wrote English note names without declaring them
Length	5:02	4:37
Notes the law committed	1,720	1,924

No note was changed in either file. A third model, minimax-m3, returned no score in two attempts.

LilyPond 2.24.4 turns each .ly into the .mid the law reads. The project used the LilyPond project’s mingw-x86_64 build, and rendering again gives the same bytes.

Admission

Both arrangements are admitted as engravings by this project under licence predicate version 4, and both are dedicated CC0 1.0. Each .ly states CC0 1.0 in its copyright field, and each .mid states no licence at all. Each receipt records the evidence for the tune, the words and the edition, the call that produced the arrangement, and the change the project made to it.

Receipt	Digest
glm-5.3	ee5a82dfc077829c0303daef54c1039d6d7b3149b0f768cdf429ee8d260569d8
kimi-k3	d044ffa49ce26ae3268e8a743ba737dd3d05e93fa7a9106afd203d100e9e4793

The recordings

Each recording is rendered through the law, not straight from the MIDI file. The law ingests the score and commits its notes, and the host plays what the law committed on the Salamander Grand Piano V3. No click sounds in the recording, because an exemplar is a score with no take.

cargo run -p host --release --locked -- fetch-piano
cargo run -p host --release --locked -- render battle-hymn-glm-5.3.wav --piece battle-hymn-glm-5.3 --voice piano
cargo run -p host --release --locked -- render battle-hymn-kimi-k3.wav --piece battle-hymn-kimi-k3 --voice piano

WAV (48 kHz, stereo, 32-bit float)	Bytes	SHA-256
battle-hymn-glm-5.3.wav	115,998,136	85b2d56723cacc6c87895a2784d2cd034dea1142e62de6d8a1a9d2867dad2eed
battle-hymn-kimi-k3.wav	106,486,216	bf49d310de5e94975b277cd1e7c57134388a0232d53bb059f12e7727baeba017

• Clean renders: no note was late or dropped, and nothing clipped.
• Reproducible: a second render gives identical bytes. CI’s piano job rendered each exemplar whole on Linux, and both SHA-256s equal the release: the same on Windows and Linux.
• Credited: each WAV carries the piano’s credit in its LIST/INFO chunk.

The page plays compressed copies made with ffmpeg 7.1: MP3 at 192 kb/s and Opus at 128 kb/s. They were encoded with its bit-exact flags, so they too come out the same every time. The uncompressed WAVs are assets of the 0.1.0 release.

The comparison

Both arrangements follow one form: an introduction, five verses each with its chorus, and a coda. The landing page sets them side by side on one time axis, section by section.

Finding the sections. tools/site/sections.py works out where each section begins:

1. It compiles a copy of each .ly with a MIDI program change placed where each section starts.
2. It reads those changes back.
3. It converts their positions to the law’s sample time through the tempo map. The tempo map is first checked against every beat the law committed.
4. It snaps every start to a note onset the law committed, and stops if there is none within two samples.

That check caught a real case. kimi-k3’s coda changes tempo on an eighth-note pickup, and interpolating between the law’s beats would have misplaced the start by about 800 samples.

The measurements. Within each section, the page measures each arrangement’s:

• length;
• notes a second;
• register, as the median note;
• average velocity;
• most notes sounding at once.

It then states the largest of those five differences in words. The table also shows each section’s range, which is not one of the five. One example: each of glm-5.3’s verses runs 28.4 seconds, and each of kimi-k3’s runs 25.7.

What is not checked

Nobody has yet done a listening test, or compared the arrangements note by note with the 1862 edition. The brief asked each model to follow the edition, and the receipts record exactly what each model was given, but neither arrangement has been proven faithful to it.