CURRENCY_SWEEP_2026-08.md

music/CURRENCY_SWEEP_2026-08.md

CURRENCY SWEEP — the music stack against the real August-2026 field

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.
Canon served: melody-first with hummable thirty-year leitmotifs; composed-never-prompted; the
no-human-composer-hire ruling (2026-08-08); the CVD §17.1 care line and the §17 floor; local-first
on the 5090; proof-before-spend.
If this document disagrees with canon, CANON WINS and this document is the defect.

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

changes no spine or registry row, installs nothing and spends nothing.

Written 2026-08-09. Lane: the August-2026 currency check over

docs/proposals/music/MUSIC_ARCHITECTURE_DECISION_BRIEF.md (revision 2, same day) and its two

step-back legs. Consumer: the AIVA benchmark session — so that the benchmark crowns an engine

against the real field and not last quarter's.

0. Why this document exists

Josh's standing floor law, invoked verbatim this sitting:

we're using the best methods and structure"

His AIVA suggestion is the FLOOR. This sweep's only job was to check whether anything better exists

right now, by fetching primary sources rather than trusting the sibling docs' candidate set.

**The headline, stated before the evidence: the floor held, and it held for a reason the sibling

docs did not have.** The commercial field moved a long way between the brief's sources and today, and

it moved *away* from us — the two engines that got better got walled, watermarked or restricted to

text prompts. **AIVA is still the only commercial engine in the field that will take a specific

melody and hand back a score.** Josh's instinct was right, and the brief undersold why.

The open field moved too, and there it moved *toward* us: three components the brief does not have

are live, licence-clean, small enough for the box, and aimed at exactly the two clauses that are

still unanswered.

---

1. Honesty register

Every claim below carries one of two marks, matching STEPBACK_OPEN_STACK.md §1.

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. Nothing here has

had its licence body pinned to docs/licence_records/; that is the step BEFORE any of it renders a

byte or takes a payment.

---

2. WHAT THE BRIEF ALREADY HAS RIGHT — survives the sweep, re-verified

Recorded so no later reader re-litigates it.

anything found this pass. Every engine checked, including the three that did not exist in the

brief's sources, decomposes into those four layers.

legal page today — VERIFIED by re-fetch of https://www.aiva.ai/legal/1.** Free/Standard are a

Non-Commercial License; Pro is a "Limited Commercial License" to monetise "on a limited set of

third-party websites: Youtube, Twitch, Tik Tok and Instagram"; a Full Copyright License exists and

reads *"Licensor hereby assigns, grants and conveys all copyrights of the MIDI and/or Audio

Composition to Licensee"*; and the training prohibition is real and tier-wide — *"Licensee is not

permitted to use the Audio and/or MIDI Composition, or any audio sample as part of a training

dataset for any Machine Learning, Deep Learning or statistical algorithm."* **The doc read the page

correctly. §3 below shows the page is not the whole story.**

CC-BY-4.0 weights, the 128-dim multihot per-frame MIDI conditioning quoted verbatim in the brief,

~71k h of mostly-instrumental stock music, 2.4B/230M. **The technical report is still forthcoming

and there are still no published evaluation metrics.** The brief's "unknown, and that is the honest

word" verdict stands at HEAD.

still 1.5 XL (4B DiT), 2026-04-02. There is no ACE-Step 2. There is still **no MIDI or melody

input of any kind**; the audio-side conditioning surface (reference audio, cover, repaint/edit,

Vocal2BGM) is exactly as the brief lists it. Code MIT; **the weights licence is still not declared

in the repo — VERIFIED absent**, which matters more now than when it was written, because the brief

leans on ACE-Step as a refiner arm.

that lineage was, if anything, understated — §3.2 below.

is untouched by anything in the field. No new component makes it stop being rung one.

pass is free, and two of the three are small enough to run alongside the existing arms.

---

3. WHAT IS STALE — named row by row

3.1 STALE ROW — STEPBACK_COMMERCIAL_ENGINES.md §3.2, the AIVA conditioning surface. It undersells AIVA, and the correction favours Josh's floor.

The doc's §3.2 describes AIVA's conditioning as preset styles, INFLUENCE ("the user uploads a MIDI

file, typically a chord progression, and AIVA analyses the harmonic structure and writes a new piece

over that structure with its own melody"), Style Designer, and explicit parameters.

**The official user manual, fetched this pass, lists four generation modes and the doc has only

one of them — VERIFIED** (https://aiva.crisp.help/en/article/general-user-manual-44klp4/):

an instrument into a full composition."*

**"Turn a specific melody into a full composition" is the composed-never-prompted primitive stated as

a shipping product feature, and no sibling document has it.** The doc's §3.2 conclusion — that

AIVA's Influence pattern is "the precedent that keeps Josh's authorship intact" — was right by

accident and for the weaker of the two available reasons. Influence is a style-transfer surface that

writes its own melody. Step by Step is the one that takes OUR tune.

Two supporting corrections from the same read, both VERIFIED:

of 60 seconds**. The doc's practitioner-sourced claim that "MIDI influences work well and audio

influences work badly" is a secondary read; the vendor documents both paths with hard minimums and

a quality note, and does not deprecate either.

automation curves**, per-track filtering, reverb, delay, Auto Staccato, section insert/delete, BPM

change, and a pattern-based percussion editor with adjustable resolution. **Dynamics automation

curves are the exact thing our own renderer does not emit** (STEPBACK_OPEN_STACK.md §5.1's "one

dynamic layer a note"). AIVA ships an expression layer as a first-class editable surface. That is

a benchmark observation, not just a feature list.

3.2 STALE ROW — STEPBACK_COMMERCIAL_ENGINES.md §4 treats Suno and Udio as a matched pair. Udio is no longer a supply route at all.

Music Business Worldwide: the new subscription platform launches "in 2026" and the existing product

runs with "creations controlled within a walled garden and the service amended in multiple ways —

including fingerprinting, filtering, and other measures."

of August 2026 the licensed platform has not shipped — LANE-REPORTED** (Billboard FAQ, RouteNote,

Chartlex, Undetectr; no primary Udio terms page fetched this pass).

Consequence for the brief: Udio is not a candidate, not a comparator and not a fallback. A model

you cannot export from cannot score a game. The §4 pairing should read "Suno, and Udio as a cautionary

tale about platform risk in this category," because the same walled-garden pressure is being applied

to Suno by the same rights-holders — LANE-REPORTED (MBW, "Universal and Suno are in a PR battle over

'walled gardens'").

3.3 STALE ROW — the same §4 has no read of Suno's own terms. They are readable, they changed, and they are better AND worse than the doc implies.

VERIFIED by direct fetch of https://suno.com/terms, Last Updated 2026-03-26:

of its right, title and interest in and to any Output owned by Suno and generated from Submissions

made by you."* Free and Basic tiers are *"solely for your lawful, internal, personal and

non-commercial purposes."*

Output or Voice Model) to create, develop or improve any competing products or services or to

power, enable or train other artificial intelligence and machine learning models, tools or

technologies."*

Two consequences, and they pull in opposite directions:

legal page's own reading. That is an uncomfortable fact and it belongs in the benchmark.

anything local. Any plan that quietly assumed "generate a corpus, train our own" is closed by both

vendors, in writing, at every tier.

"Where's V6?"; and the July-2026 Alibaba paper's leaderboard names "Suno V5.5" as the current

comparator, which is VERIFIED corroboration from a primary source).

3.4 STALE ROW — STEPBACK_OPEN_STACK.md §3.1's "exactly one open, permissively-licensed symbolic model implements the primitive we need."

The brief already corrected this from one to two (adding Composer's Assistant 2). **It is at least

three, and the third is the one built for the job we actually have.** See §4.1 — SymphonyGen.

3.5 STALE ROW — STEPBACK_OPEN_STACK.md §4's "the best conditioning surface ever built for this exact problem (JASCO) is licence-refused... anyone re-running this survey will re-discover that."

The finding is still true about JASCO. The conclusion drawn from it is now avoidable, because a

melody-conditioning adapter exists for a backbone whose licence we already accept — see §4.2,

MuseControlLite over Stable Audio Open. The open field's quality frontier and usable frontier are

closer together than the doc says.

3.6 STALE ROW — STEPBACK_COMMERCIAL_ENGINES.md §1.4 and §6, the delivery layer, cites Wwise 201 course pages and blog posts and nothing since.

Not wrong, and not structure-changing, but not current:

what's-new blog returned HTTP 403 on fetch this pass; the release exists per search indexing and

the Audiokinetic May-2026 highlights post). The doc's mechanisms — vertical layering, horizontal

re-sequencing, stingers, playlist transitions — are unchanged fundamentals.

DSP graph in the engine is a third delivery option beside Wwise and FMOD, and it is the one that

costs nothing and ships with the renderer. It belongs in §6 as a row.

into middleware**, not AI generation as a royalty-free track vending machine — LANE-REPORTED. That

is the brief's "build the delivery layer regardless of which generator wins" recommendation being

independently confirmed by practice, which is a strengthening, not a change.

3.7 NOT STALE, but now under-specified — the free render ladder has a rung the brief does not list.

STEPBACK_OPEN_STACK.md §5.2 offers exactly two sampled rungs: the CC0 incumbent and BBC SO

Discover. A third free pro-grade library exists: Orchestral Tools' Berlin Free Orchestra

LANE-REPORTED: drawn from the flagship Berlin series, 20 solo instruments and 13 ensembles, 67

articulations, four legato soloists, **round robins and multiple dynamic layers on strings and

brass, running in the free SINE player. The multiple-dynamic-layers claim is the interesting one**,

because Discover is documented as one dynamic layer per articulation on many patches — Berlin Free

may be a better test bed for the expression-emission work than Discover is.

Its licence was NOT verified this pass — the Orchestral Tools product page fetch returned

navigation only, and the EULA is at a separate /legal/license-agreement path. **That read is

outstanding and it is a blocker on using it**, exactly as the Spitfire EULA was before the brief read

it. Do not let a lane treat "widely listed as commercial-use-allowed" as clearance.

---

4. WHAT IS NEW — with evidence

4.1 SymphonyGen — the most important find in the sweep. MIT, open weights, orchestral, skeleton-conditioned, tiny.

VERIFIED by direct fetch of https://huggingface.co/SymphonyGen/SymphonyGen, with the paper at

https://arxiv.org/html/2604.25498v1:

decomposes symphonic scores along Bar, Track and Event axes with a cascading decoder.

skeleton", and the skeleton may be user-written MIDI. It also does re-orchestration** by

analysing the harmony skeleton out of an existing MIDI file. The paper is described as a

melody-aware controllable re-orchestration model that preserves melodic movement and ornament

— LANE-REPORTED for that phrasing, VERIFIED for the skeleton-conditioning mechanism.

generative component in the brief. It fits beside everything already running.

45,632 "contemporary" MIDI files.** The contemporary half carries the same unresolved question as

AMT's Lakh provenance. For a shipped commercial game that is an outstanding read, not a clearance.

Why this matters more than its size suggests. The brief's Option A stage 3 has one component,

Composer's Assistant 2, doing one job: infilling inner voices and counter-lines inside bars our code

fixed. SymphonyGen does a different job that the brief has nobody for — orchestration. It takes a

skeleton and decides which instruments carry which line, which is the layer between "we have notes"

and "we have a score worth rendering." Round 3's grade ("never any harmonies added on") and round 4's

("no variation or changes or movement") are two different missing layers, and the brief assigns both

to one model that was built for the first.

These are complements, not rivals. CA2 infills; SymphonyGen orchestrates and re-orchestrates. The

re-orchestration path is also the cheapest possible movement mechanism: the same skeleton, orchestrated

three ways, is vertical re-orchestration in the Wwise sense — produced symbolically, before any

renderer runs.

4.2 MuseControlLite — melody conditioning on a backbone whose licence we already accept.

LANE-REPORTED (arXiv 2506.18729, ICML 2025; repo https://github.com/fundwotsai2001/MuseControlLite;

the repo LICENSE file was not fetched this pass and that read is outstanding):

audio conditioning (inpainting/outpainting), combinable in any arrangement.**

control accuracy from 56.6% to 61.1% with 6.75× fewer trainable parameters than the SOTA fine-tuning

mechanism it compares against.

melody accuracy.

Why it matters: STEPBACK_OPEN_STACK.md §4.1 calls the JASCO/MusicGen-melody refusal "the

painful one" — the best melody-conditioning surface ever built, foreclosed by a CC-BY-NC weights

term. MuseControlLite is the same primitive **on the one audio backbone in the entire survey with the

cleanest declared training-data provenance** (Stable Audio Open 1.0: 472,618 Freesound + 13,874 FMA,

all CC0/CC-BY/CC-Sampling+, Audible Magic screened — the open-stack doc's own VERIFIED read). If both

licences clear, that combination is the only melody-conditioned neural audio path in the field where

we could answer a provenance challenge with a document rather than a shrug.

Caveats stated honestly. 61.1% melody control accuracy is not "it plays our tune"; it is "it

mostly follows our tune." Stable Audio Open 1.0 is capped at 47 seconds and the open-stack doc

already rates it weak on instrumental music. This is a conditioning-surface finding, not a

quality-ceiling finding.

4.3 The expressive-rendering class has two more members, and one is multi-instrument.

The brief has Symphony Rendering (ICASSP 2026) as the sole licence-clean neural realiser. Two

siblings surfaced:

for coarse expressive direction plus MIDI score for note-precise control.** Demonstrated on piano,

guitar, violin, saxophone and orchestral instruments. It reports beating **MusicGen-Melody,

Coco-Mulla and MIDI-DDSP** on FAD and CLAP, chroma similarity ~0.76–0.79, and the highest MOS across

text alignment, score alignment, expressivity and musicianship. **Weights and licence are NOT stated

in the paper — VERIFIED absent.** Demo page only.

performance synthesis. Piano only. LANE-REPORTED. Listed for completeness; the wrong instrument

family for us.

The class-level finding: "text for the feel, MIDI for the notes" is now the standard conditioning

shape in expressive rendering, and it is exactly Magenta RT 2's shape (12 MusicCoCa style tokens +

128-dim per-frame MIDI). The brief bet on the right primitive. **What it does not have is a

licence-clean member of the class with released weights** — Symphony Rendering, RenderBox and

MIDI-VALLE are all research artifacts with unstated or unfetched terms, and Magenta RT 2 is the only

one with a licence you can read and weights you can download. That materially strengthens Magenta

RT 2's position relative to the brief's own hedging.

4.4 Audio-to-audio refinement got a 2026 method — Pipeline C's missing citation.

VERIFIED by fetch of https://arxiv.org/pdf/2606.18968 — *Audio-to-Audio via Diffusion Warm

Initialization*, Cristóbal Andrade and Sebastian J. Schlecht, 2026-06-18. Instead of starting the

reverse diffusion from pure noise, it initialises from a partially denoised version of the input

audio, accelerating convergence while holding quality; demonstrated on timbre transfer and audio

enhancement.

STEPBACK_OPEN_STACK.md §9 PIPELINE C ("composed score → rough render → audio-to-audio at a swept

strength") is exactly this technique, and it now has a 2026 primary citation and a named parameter

regime instead of only the init_noise_level sweep. **This does not change the pipeline; it means the

sweep has a literature to sweep against instead of starting blind.**

4.5 New commercial entrants, and every one of them fails the disqualifying column.

This is the sweep's most decision-relevant negative result. **Four engines that did not exist in the

sibling docs' sources were checked against the one question that can disqualify: can it be given OUR

tune?**

updated March 2026; Lyria 3 Pro on Vertex AI from 2026-03-25 per Google Cloud's blog —

LANE-REPORTED). Inputs: "Text". Outputs: "Audio (music), text (lyrics)". Latent diffusion on

temporal audio latents. No audio, melody or MIDI conditioning documented at all. Every output

carries SynthID watermarking — VERIFIED, and that is a second, separate consideration for a

shipped soundtrack that no sibling doc has had to weigh. Lyria 3 Pro reaches three-minute

compositions with structural control (intro/verse/chorus/bridge) — LANE-REPORTED. **DISQUALIFIED on

the melody column.**

Inputs: **text prompt, plus an Audio Reference upload of roughly 30 seconds max "to guide the style

and sound." No MIDI, no melody file, no composition-plan input documented.** 3 s to 5 min output,

MP3 and WAV. The vendor's own claim: music *"cleared for nearly all commercial uses, from film and

television to podcasts and social media videos, and from advertisements to gaming."* The Music

Marketplace (2026-03-19 — VERIFIED) names game developers explicitly and sells three licence tiers

(Social Media, Paid Marketing, Offline). Training data is licensed — deals with Merlin and

Kobalt with revenue sharing — LANE-REPORTED. **The music-terms page (last updated 2026-05-26 —

VERIFIED) defers the actual grant to a separate "Eleven Music Model-Specific Terms" document that

was not fetched this pass. DISQUALIFIED on the melody column — but it is the licence-cleanest

commercial engine found, and if a bed-tier need ever outranks authorship it is the one to look at.**

full commercial rights and API access; MiniMax Music 2.6 released 2026-04-10 with WAV export and an

AI Cover mode. Both are text/lyrics-and-tags engines. DISQUALIFIED on the melody column on the

evidence available; neither vendor's terms were fetched.

(arXiv 2607.20253v2, Alibaba Token Foundry, 2026-07-23). Their 8B hierarchical-planner +

8B flow-matching-renderer system scores Elo 1,129, official rank range 2–3 on the Artificial

Analysis "Music with Vocals" leaderboard, and names its peers: **Suno V5.5, Mureka V8, Lyria 3 Pro,

MiniMax Music 2.6. No weights or code released.** Notably, its melody handling for cover

generation works by extracting melody tokens via source separation and pitch estimation from

reference audio — i.e. even the frontier system reaches our primitive through an audio round trip,

not a score port.

The finding this section produces, and it is the sweep's answer to Josh's question. Between the

sibling docs' sources and today, the commercial field consolidated around **licensed training data,

watermarking, walled gardens and text prompts**. Not one of the four new entrants accepts a supplied

melody. Udio, which was a candidate, stopped letting anyone export at all.

**AIVA remains the only commercial engine in the field that takes a specific melody and returns an

editable score.** The floor Josh named is not merely still standing — the sweep found no second

occupant of its category.

---

5. THE AMENDED COMPARISON TABLE — what the AIVA benchmark session carries

The first column is the one that can disqualify. Read it before anything else.

5.1 Commercial engines — the field AIVA is benchmarked against

EngineVersion / dateCan it be given OUR tune?Commercial grant for a shipped gameProvenance & platform riskVerdict for the benchmark
AIVA *(Josh's floor)*live; manual dated Sept 2025YES — the only YES in this table. "Step by Step" mode: *"turn a specific melody, an accompaniment, percussion rhythm or even an instrument into a full composition"* — VERIFIED. Plus From-Chord-Progression, plus MIDI Influence (≥8 bars)CONTRADICTORY — resolve before spending. See §5.3. Legal page: Pro = limited licence, four social platforms only. Pricing page: Pro = "Copyright owned by YOU", "Full monetization"Symbolic corpus, classical masters. No litigation found. Training prohibition on output — VERIFIED. Cloud dependency; no local optionBENCHMARK IT. It is the category of one. Its own renderer is not the product — MIDI export into our render chain is
SunoV5 / V5.5; terms updated 2026-03-26NO — text/lyrics/audio-upload; no score portStrong on paper — VERIFIED: Pro/Premier get *"assigns to you all of its right, title and interest in and to any Output"*. Free/Basic non-commercial60,000+ fingerprinted recordings in litigation; Warner settled, Sony/UMG continuing. Same tier-wide training prohibition — VERIFIED. V6 promised for 2026, not shippedReference only. Disqualified on melody
Udiowalled garden since Oct 2025NONONE — you cannot export. Downloads, stems and video disabled; licensed platform unlaunched as of Aug 2026 — LANE-REPORTEDUMG/Warner/Merlin/Kobalt deals; fingerprinting and filtering in the live productREMOVE from the candidate set. Carry only as the platform-risk cautionary row
Google Lyria 3 / 3 Pro3: 2026-02-18; Pro: 2026-03-25NO — VERIFIED. Input: "Text." Nothing elseGemini API / Vertex AI, paid preview. Model card does not address output ownership — VERIFIEDSynthID watermark on every output — VERIFIED. Training data undisclosedDisqualified on melody. The watermark is a second, independent reason to keep it out of a shipped soundtrack
ElevenLabs Music v2Marketplace 2026-03-19NO — text prompt + ~30 s audio style reference — VERIFIEDVendor claims cleared "from advertisements to gaming"; three marketplace licence tiers; actual grant lives in un-fetched Model-Specific TermsLicensed training data (Merlin, Kobalt, revenue share) — LANE-REPORTED. The cleanest provenance story of any commercial engine hereDisqualified on melody. Keep as the named fallback if a bed-tier need ever outranks authorship
Mureka V8/V9 · MiniMax Music 2.62026NO — LANE-REPORTEDMureka: paid plans claim full commercial rights, from ~$10/mo — LANE-REPORTEDunverifiedOut. No terms fetched, no melody port

5.2 Open components — the amended stack, with the three new rows marked NEW

ComponentJob in the stackMelody conditioningLicenceFits 32 GBStatus
Composer's Assistant 2infill inner voices, counter-lines, accompaniment in bars our code fixedEXACT — empty MIDI items say where to write; our theme is context, never a targetMIT code; PD + permissively-licensed MIDI training only — the best provenance of any generative component in any doctriviallybrief's stage-3 pick — unchanged and still right for infill
SymphonyGenNEWorchestration and re-orchestration — the layer between "we have notes" and "a score worth rendering"YES — beat-quantized multi-voice harmony skeleton, user-written MIDI accepted; re-orchestrates from existing MIDIMIT, weights downloadable — VERIFIED87M + 124M — smaller than everything else hereADD to stage 3 beside CA2. Different job, not a rival. Provenance read outstanding on the 45,632 "contemporary" MIDI files
Anticipatory Music Transformerinfill / accompanimentEXACTApache-2.0 code; weights licence unstated; Lakh provenance questiontriviallyunchanged — CA2 is preferred
Magenta RealTime 2neural realiser taking notes directlyDIRECT, frame-wise — 128-dim multihot MIDI + 12 style tokensApache-2.0 + CC-BY-4.0, no revenue threshold, no second chain — re-VERIFIED~5 GB at bf16unchanged. Position strengthened: it is still the only member of its class with readable terms and downloadable weights (§4.3). JAX-on-sm_120 still unproven
Stable Audio 3 Mediumaudio-to-audio refiner over our rough renderTUNABLE via init_audio + init_noise_levelStability Community (<$1M rev) + Gemma chain1.69–6.52 GBunchanged
MuseControlLiteNEWmelody-conditioned generation on a licence-clean backboneYES — melody, dynamics, rhythm and audio conditioning, combinable; 61.1% melody control accuracyrepo LICENSE not fetched — OUTSTANDING. Backbone is Stable Audio Open 1.0, the cleanest declared training provenance in the survey85M trainable adapterADD to the probe list. It is the answer to the JASCO refusal that §3.5 said had none. 47 s cap is real
ACE-Step 1.5 XLaudio-to-audio refinerindirect only — no MIDI, re-VERIFIEDMIT code; weights licence still undeclared — VERIFIED absent≥24 GB for best tierunchanged, and the undeclared weights term is now a named debt
RenderBoxNEWexpressive performance rendering, multi-instrumentYES — text (coarse) + MIDI score (note-precise)not stated in the paper — VERIFIED absent. Demo page onlyunknownWATCH, do not plan on it. Beats MusicGen-Melody / Coco-Mulla / MIDI-DDSP on FAD, CLAP and MOS. If weights land under a usable licence it outranks Symphony Rendering
Audio-to-Audio via Diffusion Warm InitNEWthe method behind Pipeline C's strength sweepn/a (technique)paper, 2026-06-18n/aCITE IT. Pipeline C now has a literature instead of a blind sweep
sfizz + VSCO 2 CE / VCSLfast preview rendererEXACTCC0, unconditional0unchanged — preview only, never final
Spitfire BBC SO Discoverfree pro-grade sampled orchestraEXACTEULA read; grant fires; §10 forbids AI training0unchanged
Berlin Free OrchestraNEWsecond free pro-grade rungEXACTEULA NOT read — BLOCKER. /legal/license-agreement fetch outstanding0Round robins and multiple dynamic layers on strings and brass (LANE-REPORTED) make it the better test bed for the expression-emission work than Discover. Read the EULA before a byte renders

5.3 The benchmark's first work item, before any track is generated

AIVA's own two pages contradict each other, and both were fetched this pass.

*"on a limited set of third-party websites: Youtube, Twitch, Tik Tok and Instagram."* A game is not

one of them. The Full Copyright License exists as a separate grant. VERIFIED.

YOU" and "Full monetization"** without restriction, 300 downloads/month, all file formats

including WAV. Free and Standard both state "Copyright owned by AIVA". VERIFIED.

Free, Standard Monthly, Standard Annually, Pro Monthly and Pro Annually plans"*, while an

**Enterprise is defined as "a business with 3 or more employees AND that generated more than $300k

of revenues in the past year"VERIFIED.** A solo developer does not obviously qualify for the

tier STEPBACK_COMMERCIAL_ENGINES.md §3.6 concluded he would need.

**So the sibling doc's conclusion — "shipping AIVA output in a commercial game requires the

Enterprise tier" — is the pessimistic reading of a genuine ambiguity, not a settled fact, and the

optimistic reading may be the vendor's actual intent.** Our own licence discipline

(fetch_sfz_stack.py: store the body verbatim, never the summary) says this is resolved by **one

email to contact@aiva.ai asking for the Pro-tier game-shipping grant in writing**, and the reply

pinned to docs/licence_records/aiva/ alongside both page bodies. **That costs nothing, it is not a

purchase, and it is the cheapest possible de-risking of the benchmark.** Do it before the benchmark

generates its first track, so the result is actionable whichever way it lands.

---

6. CHANGE TO THE RECOMMENDED STRUCTURE

The brief's structure survives the sweep. The four-layer frame, Option A, the proof-cost ordering

with the symbolic-infill proof pulled into wave one, the zero-capex DawDreamer host, the

timbre_critic wiring, the delivery layer built regardless of generator, and the CC0-sfz chain

retired to preview duty — none of it is contradicted by anything found this pass, and two parts of it

are independently corroborated (the delivery-layer recommendation by 2026 industry practice; Magenta

RT 2's position by the discovery that its whole class lacks a licence-clean alternative).

Three amendments WITHIN that structure, none of which reorder the stages:

and re-orchestrates**, and that is a layer the brief has nobody for. It is MIT, its weights are

downloadable, and it is the smallest model in the stack. Its re-orchestration path is also the

cheapest movement mechanism available — one skeleton, three orchestrations, which is vertical

re-orchestration produced symbolically before any renderer runs.

licence-refused and that is that" is no longer the end of the sentence.

bed for the expression-emission work than BBC SO Discover, because it is reported to ship round

robins and multiple dynamic layers — the exact thing our renderer is not currently exercising. **Its

EULA has not been read and that read gates its use.**

And one item that is not an amendment to the structure but is the sweep's operational finding:

resolve the AIVA licence contradiction in writing before the benchmark runs (§5.3).

---

7. HONEST LIMITS OF THIS DOCUMENT

claim is somebody else's metric.

LICENSE, the Orchestral Tools Berlin Free Orchestra EULA, the ElevenLabs "Eleven Music

Model-Specific Terms", and the AIVA Pro-tier game-shipping clarification. **None of these components

may render a byte or take a payment before its body is pinned to docs/licence_records/.**

45,632 "contemporary" MIDI files of unstated origin. For a shipped commercial game that is a read,

not a clearance.

documented input surfaces; an undocumented API parameter could exist. The documented surface is what

we can plan against.

therefore LANE-REPORTED. It does not affect the recommendation, which turns on mechanisms that have

not changed since the doc's sources.

from a primary research paper that reports its own placement on it. Treat rank ordering as

directional.

---

8. SOURCES

Fetched directly this pass (2026-08-09) — VERIFIED:

Search-indexed only — LANE-REPORTED:

Our own tree, read at HEAD:

---

9. DERIVATION

Ordered by Josh's standing floor law (2026-08-01, cross-project): his suggestions are the floor and

the loop is required to expand, check and find the gaps. The AIVA suggestion was treated as the floor

and tested against the field rather than accepted or replaced. Every candidate was asked the one

question STEPBACK_OPEN_STACK.md §0 makes the first disqualifying column — *can it be given OUR

tune?* — because that is the composed-never-prompted ruling stated as an engineering requirement.

Licence reads are face-value per memory license-reads-not-conservative-solo-dev-tooling, with

explicit triggered prohibitions named and nothing else inferred. Provenance is treated as a CVD §17.1

care question as well as a legal one. Nothing here was installed, purchased or adopted, per

proof-before-spend.

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