Verifier-Guided Twelve-Tone Composition: A Generate–Verify–Repair Harness for Symbolic Music Generation

Large language models (LLMs) often struggle with the formal constraints of twelve-tone (serial) music composition. While they can generate superficially correct sequences, they frequently collapse into degenerate textures—such as sustained chords or empty bars—that satisfy rule-based proxies while failing to produce meaningful music. This paper investigates whether a structural, neuro-symbolic intervention can enforce serial rules more reliably than prompt-based methods alone.

The authors propose a generate-verify-repair-trace harness. The process begins with an LLM proposer that generates candidate musical events. These events are passed to a deterministic verifier that checks for hard constraints, such as row-form legality, vertical non-collision, and octave equivalence. If a violation is detected, a row-aware repair operator attempts to fix the event using a five-rung escalation ladder—ranging from simple timing adjustments to LLM-based rewriting. If the repair fails, the event is dropped, and the system replans the section. The entire process is recorded in a trace, allowing for transparent diagnostics of which notes were retained, repaired, or discarded.

The harness significantly improves the quality and reliability of symbolic music generation. Across 40 controlled tasks, the audited delivery yield—the percentage of successful, rule-compliant outputs—increased from 13.3% with raw generation to 48.1% using the harness. The pass rate for collision and serialization-consistency checks rose from 33.5% to 58.3%. Crucially, the system maintained low degeneracy rates even under adversarial prompting. Expert evaluation confirmed that harness-generated candidates were preferred over raw generation in terms of coherence, perceived legality, and overall musical quality.

This work demonstrates that symbolic verification can act as a powerful guardrail for generative models in domains where rules are crisp but global structure is complex. By separating process-level diagnostics from final-artefact audits, the authors provide a template for neuro-symbolic systems that can handle structured generation tasks—such as code or formal planning—where simple prompt-based self-correction is insufficient to prevent specification gaming.