Jev Chorale Lab

A one-day experiment in two parts: first, whether a decision model knows music theory, by giving Jev undergraduate SATB part-writing exercises with a controller for voice, measure, beat, pitch, octave and duration, and grading the result with rules; second, a piano mode where Jev streams chords. A live demo is linked from the README.

Partly. Its theory is lexical: it knows scale degrees, cadences, secondary dominants and enharmonics, and can match a pitch to a chord spelled beside it, but computing Roman numerals from pitches is chance level, the chordal seventh of V7 scores 0% (it answers the seventh scale degree), and it has no representation of motion between chords, so voice leading is out of reach.

Framing, not format. CSV, ABC and LilyPond give the same failure profile, but showing only the previous beat, this beat and the next — or one sentence of prose — halved the errors and turned the melody-harmonization exercises from never-finishing into 70–91% recognizable chords at a third of the cost.

Once a score or chord progression reaches a state Jev likes, the deterministic model has no memory of its own moves and repeats or oscillates forever. The README calls this its characteristic failure, which is why the harness varies strategies and framings to break loops.

Per step at low effort, Sonnet 5 makes a third of Jev's errors at roughly 50x the cost and 10x the time, and Haiku 4.5 is about Jev's equal. Given the whole exercise and minutes to think, Sonnet 5 and Opus 5 produce near-perfect and sometimes perfect harmonizations for $0.2–0.7, while Jev's niche is $0.001 and four seconds per exercise.

npm install, copy .env.example to .env with TYPESAFE_API_KEY (and ANTHROPIC_API_KEY for the Claude comparisons), then npm run dev for http://localhost:5173. Headless runs use npx tsx against experiments/lab, experiments/quiz or experiments/oneshot; the Vite dev server proxies /api to api.typesafe.ai so the key never reaches the browser.

A never-ending stream of mellow Rhodes chords, with three Jev decisions per chord — root, then quality, then bass tone — voiced and arpeggiated by code. The README says it has no scientific value to speak of and was built because it was fun.