STEPBACK_OPEN_STACK.md

music/STEPBACK_OPEN_STACK.md

STEPBACK — the open stack, read against one question: can it play OUR tune

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.

Canon served: Josh's four recorded music grades (docs/review_candidates.json), the

melody-first / thirty-year-bar direction (memory music-direction-melody-first-30-year-bar), the

composed-never-prompted rule, the no-human-composer ruling (2026-08-08 — "composer_in_loop" is the

factory's own pass ladder, never a hire), the CVD §17.1 care line, local-first on the 5090, and

proof-before-spend.

If this document disagrees with canon, CANON WINS and this document is the defect.

TIER: PROPOSAL — ENGINE, not CANON. It sets no canon, names no region content, changes no spine

or registry row, and adopts nothing. It is a component survey with a licence read and three

candidate pipelines. Nothing in it has been installed or run this pass.

Written: 2026-08-09. Lane: music open-stack step-back. Consumer: the round-5

architecture ruling, alongside STEPBACK_POSTMORTEM.md (what our own stack did) and

STEPBACK_COMMERCIAL_ENGINES.md (how the products people pay for are built). This is the third

leg: what is actually available, open, and runnable on the box.

---

DERIVATION

DERIVED FROM:

• docs/review_candidates.json :: candidates[11..13] — PASS7_FLORES_{ROAD,FALLS,ROUNDS},

fields note, comparison_card.lines[], comparison_card.musical_axes_all_three_rounds[],

grade_verbatim (round 4, graded 2026-08-09) — the grade this doc is answering

• docs/pipeline_review/tech_research/PIPE_MUSIC_MODELS_2026-08-06.md §3.1, §3.2, §3.3, §4, §5,

§12 — the owning PIPE dossier; it binds this lane (memory pipe-dossiers-bind-generation-lanes).

This doc extends it on one axis and corrects nothing in it.

• docs/proposals/music/STEPBACK_POSTMORTEM.md §0 (the survives/retires list, including *"the

CC0-sfz chain as the final realisation path"* under RETIRES)

• docs/proposals/music/STEPBACK_COMMERCIAL_ENGINES.md §0, §0.1 (the rulings block)
• harness/music_gen/sfz_palette.py :: PLAYER, SETS — the incumbent renderer stack, read at HEAD:

sfizz 1.2.3 (BSD-2-Clause) driving VSCO 2 CE, VCSL, legato_vocal_tutorial, body_percussion

(all CC0-1.0), each pinned by commit SHA

• harness/music_gen/sa3_generate.py L1-L45 — the Stable Audio 3 adapter and, critically, its

declared legal posture on audio conditioning (§8.3 below turns on the exact wording)

• harness/music_gen/fetch_sfz_stack.py L1-L28 — the licence-body-verbatim discipline this doc follows
• docs/licence_records/ — the pinned licence bodies already on disk (sfizz, vsco2_ce, vcsl,

ace_step, stable_audio_3_medium, legato_vocal, body_percussion)

• CLAUDE.md — the content bar, the decision protocol, the do-not-invent reframe
• memory license-reads-not-conservative-solo-dev-tooling — face-value licence reads; flag only an

EXPLICITLY triggered prohibition

• memory 5090-box-live-remote-stack — Milks5090, RTX 5090 32GB, compute-clean, 24/7

NOT DERIVED (authored judgment, and why it had no canon home):

• The three candidate pipelines in §9. No canon names a round-5 architecture; that is the ruling

this doc feeds, not one it may make.

• The quality-ceiling column in every table. Where a published benchmark exists it is cited; where

none exists the column says so in those words. The judgment of what a benchmark is WORTH for our

use (orchestral cue, 60-120 s, melody-bearing) is mine and is declared as such.

• §8.3's reading of sa3_generate.py's legal posture as scoped to the EXEMPLAR corpus rather than to

our own renders. The script's docstring states the posture; the scoping inference is mine, and it

is the director's to veto because a pipeline depends on it.

---

0. The one question

Round 4 was graded, verbatim (docs/review_candidates.json, PASS7_FLORES_ROAD.grade_verbatim,

2026-08-09):

• "Round 4 is even worse than round 2 and is not shippable. The melody sucks. I hate the instruments

you are using. I hate how slow the melody is and there is no variation or changes or movement

throughout the songs. Everything sucks."

Two of those clauses are about COMPOSITION and two are about REALISATION. A model survey cannot fix

composition — STEPBACK_POSTMORTEM.md already ruled that the melodies are typed pitch tuples

developed by rules, and no amount of new renderer changes that. So this survey is organised around

the half a model CAN answer, and it asks every candidate the same question:

**Given a leitmotif that Josh owns and our engine composed — a specific sequence of specific

pitches at specific times — can this component realise THAT, or does it only produce something of

its own that sounds a bit like the prompt?**

That question is not decoration. It is the composed-never-prompted ruling stated as an engineering

requirement. A model that cannot be given a melody is not a candidate at any quality level, because

adopting it means Josh stops owning the themes. **Every table below carries a MELODY CONDITIONING

column and it is the first column that can disqualify.**

0.1 The bindings, restated so no reader has to fetch them

• MELODY-FIRST, hummable leitmotifs that survive thirty years.
• COMPOSED, NEVER PROMPTED — Josh owns the themes. A generator may develop, accompany, vary and

realise a theme it is GIVEN. A generator that invents the theme is prompting.

• NO HUMAN COMPOSER HIRE (ruled 2026-08-08). "composer-in-loop" means the factory's own pass

ladder.

• LOCAL-FIRST on the RTX 5090 32GB. Cloud/API is not a stack, it is a dependency.
• PROOF BEFORE SPEND. Nothing is bought before a measured proof on free components says the

bought thing is the remaining gap.

• THE CARE LINE (CVD §17.1) and the §17 floor. Applies to training-data provenance too: a

component whose weights were learned from a living tradition's recordings is a care question, not

only a licence question.

0.2 What this doc is NOT

It is not a re-run of PIPE_MUSIC_MODELS_2026-08-06.md. That dossier surveyed the generator field

three days ago and benched three arms on this box; its findings stand and are cited, not

repeated. Its own §9 ("what this document does not answer") left the axis this doc takes:

realisation of a supplied melody, and the renderer ladder above the CC0 sfz chain. Where a fact

appears in both, this doc marks it and moves on.

It also adopts nothing, installs nothing and spends nothing. Every "candidate" in §9 ends at a first

proof step with a cost.

---

1. Honesty register

Every claim below carries one of two marks, and the mark is the point.

• VERIFIED — a primary source was fetched this pass (2026-08-09) and the claim is read out of it.

The URL is in §11.

• LANE-REPORTED — the claim comes from a search summary or a secondary page and the primary

source was NOT fetched this pass. It may be true. It is not evidence yet.

No claim below is marked VERIFIED on the strength of a search-result snippet. That distinction is

the thing fetch_sfz_stack.py's docstring calls out — *"the licence BODY of each component is stored

verbatim ... not summarised — the summary is what the critic rejected."* Nothing here has had its

licence body pinned to docs/licence_records/, which is the step BEFORE any of it renders a byte.

---

2. The hardware envelope — what 32 GB of Blackwell actually buys

The box: RTX 5090, 32 GB GDDR7, compute-clean, 24/7 (memory 5090-box-live-remote-stack).

• Blackwell is sm_120 / compute capability 12.0, and it needs CUDA 12.8+. PyTorch 2.7.0 was the

first stable release shipping pre-built cu128 wheels with sm_120 kernels. *(LANE-REPORTED —

PyTorch forums + issue 159207.)* Our own SA3 adapter already runs torch 2.7.1+cu128 on this box

(PIPE_MUSIC_MODELS §12.1), so the PyTorch half of the ecosystem is proven here.

• The JAX half is NOT proven here. Magenta RealTime 2 — the single most interesting candidate in

this survey — is JAX/MLX only. A jax[cuda12] build on sm_120 is plausible and unverified. That is

a named risk with a cheap resolution (one install night), not a blocker.

• VRAM is not the constraint any more, and that reframes the survey. The audio field got small

while we were not looking:

**The consequence: we can run a SYMBOLIC model, an AUDIO model and a SAMPLER concurrently on one

card with room left.** A pipeline does not have to choose one of the three families. That is the

single most useful hardware fact in this document, and it is what makes §9's stacked pipelines

buildable rather than aspirational.

---

3. Symbolic / MIDI generation — the family that can be GIVEN a melody

This family is structurally aligned with composed-never-prompted: it operates on notes, so a note

we authored can be pinned and the model asked to work around it. That is the whole reason to look

here first.

3.1 The finding this family produces

Exactly one open, permissively-licensed symbolic model implements the primitive we need, and it

is the Anticipatory Music Transformer: *give it the notes Josh owns, get back everything else.* Every

other member of the family either has no melody slot (NotaGen, MuseCoco, text2midi, MIDI-LLM) or is

licence-refused (MIDI-GPT).

And the second half of that finding, stated plainly: AMT's control surface is thin. It has no

tempo curve, no section map, no dynamics envelope, no instrumentation control worth the name. Which

is precisely the gap PIPE_MUSIC_MODELS §3.2 already closed by pointing at our own code — *"the

symbolic route's control does not come from a model at all — it comes from author_head_cell.py and

orchestrate.py, which already own tempo, form, lane count, arrivals and silence by construction."*

That remains true. AMT is a candidate for ONE job: **writing the inner voices, counter-lines and

accompaniment that round 3's grade said were missing** ("never any harmonies added on"), inside bars

whose tempo, length, instrumentation and dynamic shape our code already fixed.

---

4. Audio generation — the family that sounds recorded and mostly cannot be given a tune

4.1 The finding this family produces

Three of these can be handed our music. None of them can be handed our SCORE.

• ACE-Step and Stable Audio 3 take audio — so our score must first be rendered (badly) and then

refined. The melody survives because it is physically present in the input.

• Magenta RT 2 takes notes — the only audio model in the survey with a symbolic input port, and

it happens to carry the best licence.

And the field-level fact worth stating once: **the two best-scoring open models in the world right

now (Khala, and arguably LeVo 2) are both licence-refused, and the best conditioning surface ever

built for this exact problem (JASCO) is licence-refused.** The open audio field's quality frontier

and its usable frontier are not the same frontier. Anyone re-running this survey will re-discover

that; it is written here so they do not have to.

---

5. Rendering above CC0 sfz — where "I hate the instruments" actually lives

STEPBACK_POSTMORTEM.md put *"the CC0-sfz chain as the final realisation path"* under RETIRES. This

section is what replaces it. There are two ladders — sampled and neural — and they fail in opposite

directions.

5.1 The incumbent, named precisely

Read at HEAD from harness/music_gen/sfz_palette.py:

• Player: sfizz 1.2.3, BSD-2-Clause, sfizz_render.exe, offline.
• Libraries: VSCO 2 Community Edition (CC0-1.0, commit 6dd651d5), VCSL (CC0-1.0, commit

dfcf4a49), legato_vocal_tutorial (CC0-1.0), body_percussion (CC0-1.0).

• The honest limit, written by us before Josh heard round 4 (comparison_card.honest_limits_short):

*"THE PLAYERS ARE A FREE CC0 COMMUNITY SAMPLE LIBRARY played by an offline sampler, one dynamic

layer a note."*

"One dynamic layer a note" is the load-bearing clause and it is OURS, not the library's. VSCO 2

CE ships multiple velocity layers on many instruments. A renderer that plays one layer per note

produces a dead performance from ANY library, free or bought. **Which means the first rung of the

render ladder is not a library at all — it is emitting expression.** See §8.2.

5.2 The sampled ladder — free, pro-grade, and it needs a host

The host, which is the actual new component:

5.3 The neural ladder — and it is almost entirely licence-blocked

This is the class that would render a score with learned human phrasing rather than sample playback.

It is the most exciting family in the survey and the least usable.

5.4 The finding this section produces

The neural rendering class is real, it is good, and for a commercial game it is one model wide.

Every orchestral-capable neural renderer found is CC-BY-NC or research-only except Magenta RT 2;

every licence-clean one except Magenta RT 2 is piano.

Meanwhile the sampled ladder has a free rung sitting unclimbed: **a professional orchestra recorded

at Air Studios, free, with a licence that on a face-value read grants exactly what we need — and the

only thing between us and it is a plugin host in the render chain.**

---

6. Steering and conditioning — the primitives, separated from the products

Products change every eight weeks. The primitives do not, and naming them is what lets a later lane

evaluate something this survey never saw.

1. Symbolic conditioning (notes in, notes out). A supplied note stream is held fixed and the

model completes around it. *Instances:* AMT extract_instruments() + control events; MIDI-RWKV

infilling; MIDI-GPT bar/track infilling (refused). **Fidelity to our melody: EXACT — the notes are

not regenerated, they are held.**

2. Frame-wise symbolic → audio conditioning. Note state is fed to an audio model per frame.

*Instance:* Magenta RT 2's 128-dim multihot pitch vector. **Fidelity: high by construction — the

model is told which pitches sound in every frame.** The 3 = model decides state is the honest

nuance: it can be given latitude where we want latitude.

3. Chromagram / melody-audio conditioning. A melody's pitch-class energy over time steers

generation. *Instance:* MusicGen-melody, JASCO (both refused). **Fidelity: pitch-class only —

octave and rhythm can drift.**

4. Audio-to-audio / init-audio (the img2img analogue). A rough render is partially re-noised and

denoised under a text prompt. *Instances:* Stable Audio 3 init_noise_level (VERIFIED

semantics — lower preserves more); ACE-Step Reference Audio / Cover / Vocal2BGM. **Fidelity:

TUNABLE and that is the point — there is a strength band where structure survives and timbre

changes, and finding it is a sweep, not a guess.**

5. Masked inpainting / repainting over a timeline. Regenerate bars 17-24 and hold everything

else. *Instances:* Stable Audio 3 multi-region binary masks (still UNMEASURED per

PIPE_MUSIC_MODELS §12.9); ACE-Step Repaint & Edit. **This is the primitive that lets a cue be

FIXED rather than re-rolled**, and it is the closest thing in the audio family to an event grammar.

6. Style embedding from a reference. Timbre/genre is set by an embedding rather than by words.

*Instances:* Magenta RT 2's MusicCoCa audio prompt; ACE-Step reference audio; ACE-Step LoRA from a

handful of songs. **CARE + LICENCE GATE: sa3_generate.py already declares the corpus posture —

the exemplar bytes NEVER become model input. Any use of this primitive must source its reference

from audio we own.**

7. Expression automation into a sampler. CC1/CC11/velocity/keyswitch curves shape a sampled

performance. *Instances:* DawDreamer PPQN + audio-rate automation; sfizz opcodes. **The

zero-model, zero-VRAM primitive we are currently not using, and the one that "one dynamic layer a

note" names as absent.**

---

7. The licence board — one table, go / no-go

Face-value reads per memory license-reads-not-conservative-solo-dev-tooling. **Nothing here is

cleared until its body is pinned to docs/licence_records/.**

---

8. WHAT IT MEANS FOR OUR STACK

8.1 The instrument complaint has a free answer that we have not tried

Josh said *"I hate the instruments you are using."* We are using community CC0 samples. **A

professional orchestra recorded at Air Studios is available for free, under a licence that on a

face-value read grants exactly our use.** The only thing standing between the current chain and that

one is that BBC SO Discover is a VST3 plugin and sfizz plays SFZ. **DawDreamer closes that gap in

Python, offline, at zero cost and zero VRAM.** No model, no spend, no new licence risk beyond one

body to pin. This is the highest ratio of grade-clause-answered to effort in the entire document.

8.2 …and it will NOT work on its own, which must be said before it is tried

comparison_card.honest_limits_short says *"one dynamic layer a note."* That is a property of our

REALISER, not our library. Hand a world-class orchestral library a MIDI file with static velocities,

no CC1 expression curve, no CC11 dynamics, no articulation keyswitches and no legato intervals, and

it will sound like a better sampler playing a dead performance. **Rung one of the render ladder is

emitting expression from orchestrate.py — velocity shaping, CC1/CC11 curves, articulation

selection per phrase.** The library is rung two. Attempting rung two first will produce a null result

and a wrong conclusion, and that is a prediction this document is making on the record so it can be

checked.

8.3 Audio-to-audio is available to us and a lane might wrongly believe it is banned

harness/music_gen/sa3_generate.py declares, correctly and deliberately:

*"NO AUDIO EVER CONDITIONS A GENERATION. ... The corpus legal posture

(build/audio/exemplars/CORPUS.json :: legal_posture) is absolute — 'The bytes NEVER become model

input: no training, no fine-tuning, no conditioning, no derivation.'"*

**That posture is about the EXEMPLAR CORPUS — commercial recordings we do not own, held for

measurement only. It is not, and was never, a rule against conditioning on audio we rendered

ourselves from a score we composed.** Feeding SA3 or ACE-Step our own sfz render is categorically

outside the thing that posture forbids.

This inference is mine and is flagged as NOT DERIVED. It is load-bearing: candidate pipeline C

depends entirely on it, and if the director reads the posture as blanket, C dies and should die

cleanly rather than be built on a misreading. The safe implementation is explicit — a separate entry

point with a positive control asserting the init audio's provenance is build/audio/ and never

build/audio/exemplars/.

8.4 There is exactly one licence-clean neural realiser and we have not looked at it properly

PIPE_MUSIC_MODELS §5 shortlisted Magenta RT 2 third and declined to install it on evidence:

*"its own documentation gives it no form control, no published benchmark, and no NVIDIA real-time

path."* Every word of that is true and one of the three objections does not apply to us.

"No form control" is only a defect for a system that must own form. Our system already owns form.

author_head_cell.py and orchestrate.py fix tempo, section map, lane count, arrivals and silence

by construction — PIPE_MUSIC_MODELS §3.2 says so in those words. What we lack is not form; it is

SOUND. Magenta RT 2 is the only component in this survey that takes our notes directly and returns

audio under a licence with no threshold, no registration and no second chain.

The two objections that DO stand, and they are serious:

• No published benchmark, and a ~71k-hour stock-music training prior. There is a live

possibility that our leitmotif comes back sounding like production library music, which is the

precise aesthetic Josh has rejected four times.

• It emits a 48 kHz stereo MIXDOWN, not stems. Our mix policy, the lane census and half the

measurement rig operate on lanes. A mixdown-only realiser costs us those instruments — it is not

a drop-in for the sampler, it is a different architecture with a different evidence surface.

The resolution is cheap and it is a proof, not an argument: one install night (JAX on sm_120,

unproven on this box), one leitmotif, one 30-second render, judged by the blind timbre critic

(harness/music_gen/timbre_critic.py) and by motif preservation. That is a night, not a program.

8.5 The one fork this document may not decide, restated because a new candidate sharpens it

STEPBACK_POSTMORTEM.md already named it: **is a learned model PROPOSING melodic material

composition, or is it prompting under composed-never-prompted?** That fork is Josh's.

This survey sharpens it into a clean three-way, because the candidates now separate cleanly:

• Magenta RT 2 and audio-to-audio refinement propose NO melodic material. They realise notes

they are given. Compatible with composed-never-prompted under any reading.

• AMT proposes accompaniment and inner voices around a melody it is given and never alters.

Compatible under most readings; it is the "develops, does not invent" case, and it is the direct

answer to round 3's *"never any harmonies added on."*

• NotaGen proposes THEMES. Incompatible under a strict reading. It is also the strongest

classical-craft symbolic model found, so the cost of the strict reading is real and should be

named rather than absorbed silently.

No pipeline in §9 depends on the third case. All three are buildable under the strictest reading

of the rule. That is deliberate.

8.6 What did NOT change since 2026-08-06

Re-verified this pass: no new open checkpoint has landed in the three days since the PIPE dossier.

ACE-Step's lineage still stops at v1.5/XL (there is no v1.6 or 2.0). Khala, LeVo 2 and JASCO are

still licence-refused. The instrumental quality frontier is still closed-weights. **The survey did

not go stale; the axis this doc reads was simply not the axis that dossier read.**

---

9. Three candidate pipelines, under the composed-skeleton constraint

Every pipeline below starts at the same place — **a leitmotif Josh owns, a score our engine

composed** — and differs only in how that score becomes sound. None invents melody. Each ends at a

first proof step with a cost, per proof-before-spend.

They are stackable, not exclusive. A is the floor and both others sit on top of it.

---

PIPELINE A — "Better players" · composed MIDI → pro-grade sampled orchestra

orchestrate.py  ──emit MIDI + CC1/CC11/velocity/keyswitch──►  DawDreamer (GPLv3, build-time)
                                                                    │  hosts VST3
                                                                    ▼
                                                     BBC SO Discover  ──►  per-lane stems
                                                                    │
                                                        existing mix_policy.py + battery

• What it answers: *"I hate the instruments you are using"* — directly, and only that.
• Models: none. VRAM: zero. Spend: zero.
• Melody fidelity: EXACT. Nothing regenerates anything; a sampler plays the notes it is given.
• Licence surface: Spitfire EULA (face-value GO, body to pin) + GPLv3 build-time tool +

Steinberg VST3 SDK terms. No model licence at all.

• Everything downstream survives: stems, lane census, mix policy, the whole measurement rig.

This is the only pipeline with zero evidence-surface cost.

• The catch, stated as a prediction (§8.2): without the expression layer this produces a better

sampler playing the same dead performance. **Rung one is orchestrate.py emitting expression;

the library is rung two.** Build them in that order or the result is a null.

• Second catch: Spitfire's plugin authorises through their own installer, and headless VST

hosting of authorised commercial plugins is where DawDreamer's documented plugin-compatibility

caveats bite. If it does not host, REAPER CLI is the fallback and it costs money — which is

exactly the kind of spend proof-before-spend permits, because A's proof would have earned it.

• FIRST PROOF (one night, zero spend): take the round-4 Flores cue's existing MIDI, add a CC1

expression curve and velocity shaping, render it BOTH ways — sfizz/VSCO2 and DawDreamer/BBC SO

Discover — and run both through timbre_critic.py and the existing battery. **Two numbers and two

files. If the timbre critic does not move, the instrument hypothesis is wrong and we learned it in

a night.**

---

PIPELINE B — "Neural realiser" · composed MIDI → Magenta RT 2 → audio, our code owns form

orchestrate.py ──► note stream ──► 128-dim multihot pitch state per frame ─┐
                                                                            ├─► Magenta RT 2 ──► 48 kHz stereo
style reference (audio WE own, or text) ──► 12 MusicCoCa tokens ───────────┘         │
                                                                                     ▼
                                                            our seam/continuity layer (loop_seams.py)

• What it answers: *"I hate the instruments"* AND, potentially, the phrasing/liveness gap that no

sampler closes — because the character comes from a model trained on recorded performance.

• Models: Magenta RT 2 base (2.4B) or small (230M). VRAM: ~5 GB. Spend: zero.
• Melody fidelity: HIGH BY CONSTRUCTION. The model is told which pitches sound in every

frame; the 3 = model decides state is deliberate latitude we control per-note.

• Licence surface: the cleanest available. Apache-2.0 + CC-BY-4.0, no revenue threshold, no

registration, no second chain, *"Google claims no rights in outputs."*

• Why it is not simply better than A: three real costs, in order of severity.

1. Mixdown, not stems. Our mix policy, lane census and several instruments operate on lanes.

This pipeline costs us that evidence surface, or forces per-lane rendering and a re-mix — which

is untested and may not sum coherently.

2. ~20 s effective receptive field, 25 Hz frames. Long-form continuity over an 88-second cue is

OUR problem to solve by stitching. We already own loop_seams.py, so this is work, not a wall.

3. ~71k h of stock-music training prior, and no published benchmark. The genuine risk that our

leitmotif returns sounding like production library music.

• Plus one unknown: JAX on Blackwell sm_120 is unverified on this box. WSL2 Ubuntu-24.04 is

present (PIPE_MUSIC_MODELS §5), so the path exists; nobody has walked it.

• FIRST PROOF (one night, zero spend): install JAX/CUDA, render ONE 30-second leitmotif statement

through mrt2_base with a text style prompt, and score it against the same cue rendered through

Pipeline A. Judged by timbre_critic.py (blind) plus a motif-preservation check. **The install is

the risk; the render is minutes.**

---

PIPELINE C — "Refiner ladder" · composed score → rough render → audio-to-audio at a swept strength

orchestrate.py ──► MIDI ──► sfizz or Pipeline A render (structure is now physically present)
                                              │
                                              ▼   init_audio, init_noise_level swept
                        Stable Audio 3 Medium (MIT-ish/Community) ──or── ACE-Step 1.5 (MIT)
                                              │
                                              ▼
                          motif_compare.py + timbre_critic.py, scored across the sweep

• What it answers: the *"sounds like a mockup"* gap — production character, room, ensemble

cohesion, performance grit — while the composition, form, arrivals and silence survive because they

are physically present in the init audio.

• Models: SA3 Medium (~6.5 GB) or ACE-Step 1.5 XL (fits at ≥24 GB). Spend: zero — **both are

already installed and licence-pinned on this box.**

• Melody fidelity: TUNABLE, and finding the band IS the experiment. VERIFIED semantics:

*"lower values preserve more of the original."* At a low noise level you get the same dead

performance under a haze; at a high one the melody drifts away from Josh's theme. There is a band

between and it is found by a sweep scored on melody preservation, not by a guess.

• Licence surface: already pinned for both. SA3 carries the **live, unsatisfied Section III

registration obligation**; ACE-Step carries none.

• The blocking premise: §8.3 — that conditioning on our own render is outside the corpus

posture. If the director reads that posture as blanket, this pipeline does not exist. That read

is the cheapest thing in this document and should happen first.

• Two honest costs: (i) diffusion refinement smears onsets, which degrades exactly the

arrival-placement and silence-budget dimensions our cards measure — measurable with instruments we

already have; (ii) it produces a mixdown, same evidence-surface cost as B, though here we keep

the pre-refinement stems as a fallback mix.

• FIRST PROOF (one night, zero spend, zero install): take the round-4 Flores cue's existing

render, sweep init_noise_level across 0.1 / 0.2 / 0.35 / 0.5 under one caption, and plot melody

preservation against timbre-critic distance. **The output is a curve, and the curve either has a

usable band or it does not.** This is the cheapest proof in the document because every component

is already on disk.

---

9.1 How they compose, and the recommended order

They stack: **A is the renderer floor. C rides on A's output. B is an independent track that

bypasses the sampler entirely.**

RECOMMENDATION (this lane's, and the director's to overturn): run C's sweep first because it

is hours on components already on disk and it either finds a band or eliminates a whole family;

then A rung one (expression from orchestrate.py) because §8.2 predicts everything else in A

depends on it; then A rung two (BBC SO Discover through DawDreamer); then B as the

independent track once the JAX install has a night to spare.

THE STRONGEST OBJECTION to all three: none of them touches *"the melody sucks"* or *"no variation

or changes or movement."* Those are composition defects, STEPBACK_POSTMORTEM.md located them in

pass3_flores.py:114-144, and a renderer cannot fix a tune. **If round 5 ships better instruments

playing the same melody, it will be graded down again, and this document will have been part of the

reason. The only component here that addresses the composition side at all is AMT** (§3.1) —

harmonising and counter-lining a theme it never alters — and it belongs in the round-5 architecture

alongside whichever renderer wins, not after it.

---

10. What this document does not answer

• No licence body is pinned. Every GO in §7 is a face-value read of a page, not of a stored

verbatim body under docs/licence_records/. **BBC SO Discover, DawDreamer, AMT's weights and

NotaGen all need that step before a byte renders.**

• AMT's weights licence and Lakh MIDI provenance are both unread. For a shipped commercial game

the provenance question is the more serious of the two.

• Nothing was installed or run. Every number in §9's proof steps is a proposed measurement.
• The expression-layer work in orchestrate.py (§8.2) is scoped as a prediction, not designed.

It is the actual engineering in Pipeline A and it is not costed here.

• Stable Audio 3's multi-region mask — the primitive that got it shortlisted — is still

UNMEASURED (PIPE_MUSIC_MODELS §12.9) and is not exercised by any pipeline above.

• Care-tier review of training-data provenance for Magenta RT 2 and ACE-Step against CVD §17.1

has not been run. Neither vendor enumerates its corpus at the granularity that question needs.

• Adaptive/interactive scoring — Magenta RT 2's actual design purpose, and arguably relevant to a

game — is out of scope here and boarded.

---

11. SOURCES

All fetched or searched 2026-08-09.

Audio generation

• ACE-Step 1.5 — https://github.com/ace-step/ACE-Step-1.5 · README:

https://github.com/ace-step/ACE-Step-1.5/blob/main/README.md · paper:

https://arxiv.org/abs/2602.00744 · weights: https://huggingface.co/ACE-Step/Ace-Step1.5

• Stable Audio 3 — https://github.com/Stability-AI/stable-audio-3 · inference workflows:

https://github.com/Stability-AI/stable-audio-3/blob/main/docs/workflows/inference.md

• Stable Audio Open 1.0 — https://huggingface.co/stabilityai/stable-audio-open-1.0
• Stability AI Community License — https://stability.ai/license
• Magenta RealTime 2 — https://github.com/magenta/magenta-realtime · model card:

https://huggingface.co/google/magenta-realtime-2 · docs:

https://magenta.github.io/magenta-realtime/ · original announcement:

https://magenta.withgoogle.com/magenta-realtime

• MusicGen model card (weights CC-BY-NC-4.0) —

https://github.com/facebookresearch/audiocraft/blob/main/model_cards/MUSICGEN_MODEL_CARD.md

• YuE — https://github.com/multimodal-art-projection/YuE ·

https://huggingface.co/m-a-p/YuE-s1-7B-anneal-en-icl

• DiffRhythm — https://github.com/ASLP-lab/DiffRhythm
• HeartMuLa — https://github.com/HeartMuLa/heartlib
• LeVo 2 / SongGeneration — https://github.com/tencent-ailab/SongGeneration ·

https://arxiv.org/html/2606.30642v1

Symbolic / MIDI

• Anticipatory Music Transformer — https://github.com/jthickstun/anticipation ·

https://arxiv.org/abs/2306.08620 · weights: stanford-crfm/music-medium-800k

• NotaGen — https://github.com/ElectricAlexis/NotaGen ·

https://huggingface.co/ElectricAlexis/NotaGen · https://arxiv.org/abs/2502.18008

• MIDI-RWKV — https://arxiv.org/html/2506.13001v2
• MIDI-GPT — https://arxiv.org/pdf/2501.17011
• text2midi — https://arxiv.org/html/2412.16526v2

MIDI-to-audio rendering

• Multi-Aspect Conditioning for Diffusion-Based Music Synthesis —

https://benadar293.github.io/multi-aspect-conditioning/ ·

https://github.com/benadar293/multi-aspect-conditioning

• RenderBox — https://arxiv.org/abs/2502.07711 · https://renderbox-page.vercel.app/
• nii-yamagishilab midi-to-audio — https://github.com/nii-yamagishilab/midi-to-audio
• MIDI-VALLE — https://ismir2025program.ismir.net/poster_208.html
• Pianist Transformer — https://openreview.net/forum?id=Kq9MMEynGW

Renderers, hosts and libraries

• Spitfire BBC SO Discover — https://www.spitfireaudio.com/bbc-symphony-orchestra-discover · FAQ:

https://support.spitfireaudio.com/en/articles/11815834-bbc-symphony-orchestra-discover-faq ·

EULA: https://www.spitfireaudio.com/info/eula/

• DawDreamer — https://github.com/DBraun/DawDreamer · docs: https://dirt.design/DawDreamer/ ·

plugin compatibility: https://dirt.design/DawDreamer/compatibility.html · paper:

https://archives.ismir.net/ismir2021/latebreaking/000001.pdf

• REAPER CLI — https://github.com/ReaTeam/Doc/blob/master/REAPER-CLI.md · https://www.reaper.fm/
• VSCO 2 Community Edition — https://versilian-studios.com/vsco-community/
• Virtual Playing Orchestra — https://virtualplaying.com/virtual-playing-orchestra/ (REFUSED, §5.2)

Platform

• PyTorch sm_120 / Blackwell — https://github.com/pytorch/pytorch/issues/159207 ·

https://docs.salad.com/container-engine/tutorials/machine-learning/pytorch-rtx5090

Repo-internal (canon and lane-canon)

• docs/pipeline_review/tech_research/PIPE_MUSIC_MODELS_2026-08-06.md §3.1-§3.3, §4, §5, §12
• docs/proposals/music/STEPBACK_POSTMORTEM.md · docs/proposals/music/STEPBACK_COMMERCIAL_ENGINES.md
• harness/music_gen/sfz_palette.py · harness/music_gen/sa3_generate.py ·

harness/music_gen/fetch_sfz_stack.py

• docs/review_candidates.json :: PASS7_FLORES_{ROAD,FALLS,ROUNDS}

Generated by harness/site/structure_site.py — the URL path is the repo path. review root