This paper presents a neuro-symbolic harness for twelve-tone composition that wraps a language-model proposer in a generate-verify-repair-trace loop. Symbolic checks raise constraint-checked delivery from 13.3% to 48.1% across 40 controlled tasks and four paired models, while the system abstains on the remaining 51.9% of runs. A narrower collision and serialisation-consistency check improves from 33.5% to 58.3%. Degeneracy stays near 0.05, including under adversarial prompts. Five experts descriptively preferred harness outputs in adherence, perceived legality, coherence, and overall quality, although the abstract does not provide detailed statistical results.
No heat snapshots are available in the last 24 hours.