How does music evolve?
SeraWorld
SeraWorld studies musical states, their relationships, and the consequences of change.
Inside SeraWorldMUSIC / RESEARCH / PRACTICEMusic,
Music,
intelligence
& form.
Modeling how music evolves.
Creating with AI, with composers in control.
CALLING / 2026For small chamber ensemble
Musical intelligence · Human–AI collaborationSCROLL TO DISCOVER
01 / RESEARCHUpdated
Two questions.
One musical world.
My research focuses on intelligent musical systems that model how music evolves and collaborate reliably with human composers over long creative processes.
How should AI act on music?
SeraForge
SeraForge studies editing tools, project memory, and reversible workflows that preserve a composer's intent.
Inside SeraForgeA shared foundation: Sera Core
Authoritative score state, musical operations, and reversible patches.
02 / SELECTED WORKSystems for thinking through music
Ideas, made
tangible.
Research takes shape through instruments, software, and experiments. A selection from the Sera ecosystem.
SERAFORGE / OCTOBER 2026Working research prototype
A composer's intent.
A reviewable musical action.
Plan with an agent, develop material through tools, then inspect and hear the draft before accepting it. The score remains under the composer's control.
- 01Describe & plan
Set the musical goal, retrieve project memory and protect what should stay.
- 02Develop & inspect
Work from actual source notes, compare changes and repair within explicit scope.
- 03Preview & decide
Review the score, listen, revise or accept. Keep a reversible history.
Material with an identity
Inspect exact phrase relationships, including transposition, inversion and rhythmic scaling.
Repair with visible tradeoffs
Protect rhythm and voices while inspecting pitch and register candidates.
Optional musical intelligence
Use World candidates through public interfaces, with Forge checks and acceptance still required.
Implemented tools and a locally verified Windows build. Long-horizon musical quality and comparative human preference remain research questions.
COMPOSITION / 2026
CALLING
For small chamber ensemble
03 / SOFTWARE IN ACTIONSeraEdit · Interactive replay
A musical idea.
A traceable change.
Follow a real edit from an instruction to a validated score revision. Explore six recorded examples from the offline software demonstration.
Recorded software execution. The walkthrough replays saved results; it does not call an AI model.
Download evidence ↗Precise musical scope
Address measures, staves, voices and individual events. Bind every proposal to the score it was created for.
SERAEDIT / SERA COREIntent that persists
Retain project memory and lock musical properties, such as a motif's rhythm, across an editing session.
SERAFORGEChanges you can inspect
Review differences, validation results and candidate alternatives before accepting a reversible change.
SERAFORGE / SERA COREMusical consequences
Study how timing, voice and context affect the predicted consequences of an edit, including when to abstain.
SERAWORLD / IN DEVELOPMENT04 / RESEARCH & RELEASESEvidence, questions, next steps
From exploration
to evidence.
An accepted SoftwareX paper, a working composition research environment, and new tools for making musical decisions inspectable.
ACCEPTED PAPER
SoftwareX
Accepted for publicationEditorial decision
Yuan Gao · SoftwareX · Accepted 2026
SeraEdit: Scoped and Validated Language-Guided MusicXML Editing
A local-first editing system that turns language-guided proposals into bounded, validated changes, with explicit protection for the rest of the score.
20Synthetic benchmark scores
120Editing tasks in the release
6 / 6Offline demo cases passed
Rerun September 15, 2026
Rerun September 15, 2026
- Original software archive · v1.0.0
- Submitted to SoftwareX
- Minor-revision decision received
- Real-provider comparison completed · 360 initial runs
- R1 source and evidence archive published
- Accepted for publication in SoftwareX
Acceptance confirmed for manuscript SOFTX-D-26-01135R1; publisher production follows. The links below identify software and reproducibility archives. The online article DOI and final citation will be added when available.
Protection has
a coverage cost.
On 120 development tasks, three workflows completed 360 initial DeepSeek runs. The full Sera workflow preserved non-target content in its produced editing scores, but refused more executable requests.
| Workflow | Strict edits | Leaking scores |
|---|---|---|
| Full Rewrite | 72 / 110 | 17 / 105 |
| Patch Only | 65 / 110 | 18 / 92 |
| Sera Full | 63 / 110 | 0 / 73 |
Zero measured leakage among produced scores does not mean every requested edit succeeds.
Study scope & evidence
Twenty synthetic source scores; 110 executable edits and 10 expected refusals per condition. Leakage denominators count produced editing scores. Sera Full made 34 erroneous edit refusals and received 43 repair calls. The 360 metrics recomputed exactly; 14 integrity checks passed. This is a development system comparison, with overlapping pilots and unequal repair budgets, not a held-out superiority or musical-preference study.
From musical claims
to actual notes.
Source-grounded thematic planning and complete phrase comparison make exact material relationships inspectable. Bounded repair tools search pitch and register choices while preserving protected musical properties.
A repair can reduce voice-leading findings while increasing melodic motion. The composer still decides whether it works as music.
Study scope & evidence
October 18, 2026 · A retained unfinished draft had 33 ensemble findings. With a 7-semitone bound on new or worsened leaps, finite octave search left 31; with a 12-semitone bound, it left 3, while measured inner-voice motion rose from 178 to 382 semitones. These were analyst-configured engineering probes on 16 written bars of a 64-bar framework, with the main melody and bass protected and no accepted score commit. They are not autonomous-model or musical-preference results.
The frozen local engineering record reports 2,514 passing regressions, a packaged Windows build and acceptance/undo checks. This website session checked those records without rerunning the Forge suite. V0.3 research acceptance and complete long-form musical quality remain open.
More intelligence.
The same human control.
Forge can now request rule-based or local neural candidates through public World interfaces, inspect their source material and route selected proposals into its existing draft workflow.
Forge continues to run without World. Every candidate still requires validation, preview and explicit human acceptance.
Integration scope & evidence
October 17–18, 2026 · Recorded real-model calls reached the optional World tools. One local frozen-model run returned a 16-bar, six-instrument candidate; all 166 sounding notes matched in pitch, onset, duration and part after import. Preview dynamics use coarse categories. Later workflow traces reused model-selected arguments with test-driver orchestration, so they do not establish full autonomous creation.
Complete 64-bar development, matched general-LLM versus World-assisted listening, human preference and a causal quality gain remain unverified. Providers return suggestions; they cannot commit authoritative score state.
Knowing when
a prediction is uncertain.
The SW-002 prototype models editing consequences with event timing and voice context. Its latest recorded contextual-calibration test passed 4 of 12 seed-by-work reliability units.
Musical state→Proposed action→Consequence + uncertainty
Broad intervals and rejected actions still limit useful coverage. The strict reliability gate has not passed.
Study scope & evidence
September 10, 2026 · 10 new calibration works and 4 development-validation works passed source admission. Independence refers to this predictor's training and earlier calibration, not all historical project data. No newer calibration success is recorded here. Multi-step planning and musical usefulness remain open.
A playable file.
An unresolved musical result.
Six fixed original-score excerpts now sit alongside three generation routes. All 18 generated excerpts render through official MuseScore playback after a notation repair that preserves sounding events.
The composer still rejects the generated music. G0 has not passed; the recursive G1 experiment stays paused.
Study scope & evidence
October 2, 2026 · These are development diagnostics in a separate experiment workspace, not new SW-002 calibration results or Forge editing outcomes. The earlier coarse representation lost short notes, polyphony and source tempo. Originals retain source tempo; generated excerpts use 100 BPM and simplified prefixes, so this is not a matched or blind listening study. Rendering repairs do not fix learned musical development or prove recursive collapse.
ONE ECOSYSTEM. DISTINCT QUESTIONS.
SeraWorld models musical evolution. SeraForge transforms that intelligence into reliable human–AI musical action. Both share authoritative score infrastructure through Sera Core; Forge runs without World. Modeling tests, editing benchmarks and human studies remain separate.
05 / RESEARCH NOTESA record of work in progress
An evolving
practice.
Where the work stands, what remains open, and the questions that guide the next experiment.
Research snapshot · . Experiment dates are stated individually. SeraEdit is accepted for publication; development measurements and composer feedback retain their own evidence limits.
06 / ABOUT
Yuan Gao.
AT THE INTERSECTION OF MUSIC & AII study how intelligent systems can model musical evolution and collaborate with composers.
Through the Sera research ecosystem, I connect musical representation and prediction with tools, memory, validation and reversible creative workflows.
My SeraEdit paper was accepted for publication in SoftwareX in October 2026. Current work on SeraForge asks how AI can participate reliably in longer composition and editing processes while composers retain creative control.
Follow the research07 / CONTACTResearch · Software · Composition
Let's start
a conversation.
For research conversations, software questions, or musical collaborations, you can reach me at any of these addresses.