# INTEGRATION-cycle-4 — AMT "what's NEW since cycle-3" sweep > **Status:** This is the **persisted delta** for the 2026-08-07 cycle-4 sweep. > Cycle-3 covered the MuScriptor hot-take + MIROS/HeartMuLa/mason369/Magenta RT2 (last commit Jul 21 2026 on `muscriptor/muscriptor`). > Cycle-4 covers the **17-day post-MuScriptor window (Jul 21 → Aug 7 2026)** + a 6–12-month backfill for 2026 releases cycle-3 missed. ## 1. TL;DR - **Cycle-3's recommendation still holds.** MuScriptor is the SOTA on full-mix multi-instrument; Basic Pitch is still the per-stem fallback; HeartMuLa is still the Apache-2.0 lyrics path. No cycle-4 candidate beats the cycle-3 AMT pipeline on quality. - **Four new watchlist entries** extend (don't replace) the existing pipeline: - **ACE-Step v1.5** (Apache 2.0) → replaces YuE as the **budget music-gen default** (8 GB VRAM, 15× faster) - **Libretto** (UC Berkeley) → the first **symbolic AMT agent interface** (closes the cycle-2c "no good AMT-agent shape" gap) - **Rubato** (arXiv:2605.24291) → the first **one-pass audio→timestamped score** system (closes the "cascaded hack" gap) - **Stable Audio 3.0** (May 20 2026) → the only **commercial-safe** open music-gen (Stability Community License + AudioSparx/Freesound CC data) - **The demo at `demo/` is refreshed to cycle-4.** The existing Long Horizon Prompt tool already meets the spec (6 formats, copy buttons, vanilla JS, no server). Part A is updated to surface the 4 new cycle-4 entries alongside the cycle-3 cards. No code rewrites. - **No removals from the watchlist.** basic-pitch stays dormant-flagged. omnizart stays on watch. MIROS public weights still gated. MuScriptor PyPI still forthcoming. ## 2. What changed in the pipeline (cycle-4) ``` ┌────────────────────────────────────────────────────────────────────────────┐ │ MUSIC GENERATION (lyrics→song) — cycle-4 changes │ │ │ │ Cycle-3 default for full-quality: YuE (Apache 2.0, 16 GB VRAM, ~5 min) │ │ │ │ Cycle-4 default for full-quality: YuE unchanged │ │ — Same license, same quality, more time on GPU │ │ │ │ Cycle-4 default for budget: ACE-Step v1 (NEW) ⭐ │ │ — Apache 2.0, 8 GB VRAM, ~2 sec/song on A100 (15× faster) │ │ — 459M / 3.5B variants, fits in any hobbyist rig │ │ — Apache-2.0 → license posture identical to YuE │ │ — Replaces the cycle-3 "YuE needs 16 GB" gap for low-VRAM users │ │ │ │ Cycle-4 add for commercial-safe: Stable Audio 3.0 Medium (NEW) ⚠ │ │ — Stability Community License (free for orgs < $1 M ARR) │ │ — Trained on licensed AudioSparx + Freesound CC │ │ — 6:20 output, 1.4B params │ │ — NOT Apache-2.0; commercial use above ARR threshold needs license │ │ — Watchlist, not default — license posture is the constraint │ │ │ │ WHY: ACE-Step matches YuE's Apache-2.0 posture while running on │ │ ⅛ the VRAM and finishing 15× faster. Same audio quality bar for the │ │ hobbyist user. YuE retained for users with 16 GB+ who want the longest │ │ output (3-4 min tracks at 44.1 kHz). │ └────────────────────────────────────────────────────────────────────────────┘ ┌────────────────────────────────────────────────────────────────────────────┐ │ AGENT BRAIN (symbolic) — cycle-4 changes │ │ │ │ Cycle-3 LALM (audio side): nvidia/audio-flamingo-3-hf (Apache 2.0) │ │ Cycle-4 LALM (symbolic): Libretto (UC Berkeley, arXiv:2606.22708) NEW │ │ — LLM-native grammar with onset slots, voices, bar-level organization │ │ — Corpus-calibrated statistical music cloud │ │ — Musician-readable feedback for retrieval / generation / revision │ │ — Closes the cycle-2c "no good AMT agent interface" open question │ │ — License: check on first commit (paper only at this stage) │ │ │ │ WHY: Audio-Flamingo-3 is an audio-in / text-out model. It's great at │ │ "what is this audio?" but bad at "rewrite this score to fix the bridge". │ │ Libretto is text-in / text-out on a symbolic-music grammar — the right │ │ shape for agentic music editing. The two together cover both sides. │ └────────────────────────────────────────────────────────────────────────────┘ ┌────────────────────────────────────────────────────────────────────────────┐ │ END-TO-END AUDIO→SCORE — cycle-4 changes │ │ │ │ Cycle-3 path (cascade): │ │ audio → AMT (MuScriptor) → MIDI → music21 cleanup → Verovio engrave │ │ │ │ Cycle-4 path (single-pass) — WATCHLIST: │ │ audio → Rubato → timestamped notation (one pass) │ │ — Beats cascades on notational accuracy │ │ — New "InterMo" text representation for sequence-to-sequence training │ │ — Demo: https://nctamer.github.io/rubato-transcription │ │ — Code release pending (check before adopting) │ │ — License: research-only at paper stage │ │ │ │ WHY: the cascade path introduces notation errors at every stage boundary. │ │ Rubato's prompt-conditioned encoder-decoder predicts time + pitch + │ │ notation in one autoregressive pass. The right answer for the "I just │ │ want the sheet music" use case — but adoption depends on code release. │ └────────────────────────────────────────────────────────────────────────────┘ ``` ## 3. What changed in the demo (cycle-4) The existing `demo/` directory from cycle-3 already satisfies the Long Horizon Prompt spec: - 6 production-ready output formats (horizontal / vertical / square video, audio, music production, social post) - Copy button on every output card - "Copy all 6" bundled copy at the bottom of the input column - Vanilla JavaScript, no server, no API calls, no rate limits - Dark + light theme via the existing design system - Cmd/Ctrl+Enter to regenerate, optional keyword list, tone/audience/voice/length dropdowns **Cycle-4 changes are minimal — extend, don't rewrite.** Two edits to `demo/index.html`: 1. **Hero** — change cycle-3 → cycle-4 and add a one-line "what's new" subtitle pointing at the cycle-4 entries 2. **Part A** — replace the 5 cycle-3 cards with a 5-card cycle-4 set that **rotates the cycle-3 winners to a compact "carried forward" line at the top of Part A** and surfaces the 4 new cycle-4 entries (ACE-Step, Libretto, Rubato, Stable Audio 3.0) as fresh cards No changes to Part B (Long Horizon Prompt). The 6 output cards, the templates, and the copy buttons are unchanged. The demo's `README.md` gets one extra section (the cycle-4 delta). The `styles.css` and `app.js` are unchanged. **File map after cycle-4:** ``` demo/ ├── index.html # refreshed Part A (cycle-4 entries) + cycle-3 cycle-4 hero subtitle ├── styles.css # UNCHANGED ├── app.js # UNCHANGED (the LH prompt is cycle-3 code; no JS edits needed) └── README.md # added §"cycle-4 delta" entry ``` **No code added in cycle-4.** This is a *card-rotation* cycle, not a *new-tool* cycle. The Long Horizon Prompt was the cycle-3 new tool; cycle-4 is a sweep + a card refresh. ## 4. What to do next (cycle-5 watchlist) The cycle-4 sweep surfaced 6 items worth re-checking in cycle-5: 1. **MIROS public weights** — the canonical artifact is still the closed AI4Musicians challenge repo. If they post a public checkpoint in the next 30 days, drop it into `RESOURCES.md` and update the eval pipeline. 2. **MuScriptor PyPI release** — the model card still says "PyPI release forthcoming" as of Aug 7 2026. When the `pip install muscriptor` package ships, update `setup/pipeline_starter.py` and `trusted_packages.txt`. 3. **Rubato code release** — the paper (arXiv:2605.24291, May 22 2026) is the only artifact. If the code lands, evaluate against the cycle-3 MuScriptor → music21 → Verovio cascade on MAESTRO v3.0.0 test set. 4. **Libretto code release** — paper (arXiv:2606.22708, Jun 21 2026). When code ships, test the symbolic-side agent against the cycle-3 Audio-Flamingo-3 audio-side agent on a "describe this piece + suggest edits" task. 5. **ACE-Step v1.5 XL (4B DiT)** — the XL line released Apr 2 2026 is the high-quality music-gen option. Re-benchmark when the `turbo-rl` (RL post-trained) variant lands. 6. **V2N (ISMIR 2026)** — first video-only piano transcription. Long-term watch; not in the default pipeline but worth tracking for "self-driving lab" + "see the hands" use cases. **Nothing to drop.** The watchlist removals from cycle-3 still hold. ## 5. Where to look (file map) | What | Where | |---|---| | Full categorized list (this cycle) | `dashboard/RESOURCES-cycle-4.md` | | Carried forward (cycle-3) | `dashboard/RESOURCES-cycle-3.md` | | Carried forward (cycle-2c) | `dashboard/RESOURCES-delta-cycle-2c.md` | | Carried forward (audio sweep) | `dashboard/RESOURCES-audio.md` | | This cycle's demo refresh | `dashboard/demo/index.html` (Part A only) | | This cycle's demo README delta | `dashboard/demo/README.md` | | Cycle-3 demo README (the LH prompt spec) | `dashboard/demo/README.md` §"Part B" | | Cycle-3 hub card | `dashboard/index.html` (the "AMT Demo + Long Horizon Prompt" card) | | Pipeline stage config (to extend) | `setup/pipeline_starter.py` (unchanged this cycle) | | Trusted packages list (to extend) | `trusted_packages.txt` (no new entries this cycle) |