docs: document Stable Audio 3 setup

Author: emeisazam

README.md

@@ -13,6 +13,13 @@ Current release: `0.2.0`.
 GERM is an independent Sonic Field Labs project. It can use Stable Audio 3
 providers, but it is not an official Stability AI product.
 
+The Listening Stack remains model-agnostic at its contracts. GERM's first
+fully developed local generation path is more specific: it is designed around
+Stable Audio 3 and uses its text-to-audio, audio-to-audio, inpainting,
+continuation, variable-length, and LoRA primitives as material for cultivation.
+The upstream code is MIT; the currently released local weights are gated and
+carry separate Stability AI and component terms.
+
 ![GERM dashboard with a prompt module in the cultivation graph](docs/assets/germ-dashboard.png)
 
 ## Try It Without a Model
@@ -47,8 +54,9 @@ checks but not a creative quality demonstration.
 | Path | Operating system | Hardware and software |
 | --- | --- | --- |
 | Mock service | macOS or Linux | Python 3.10+, `uv`; CPU only; no weights |
-| Stable Audio 3 MLX | Apple Silicon macOS | Official `sa3` MLX tools and separately accepted model weights |
-| Stable Audio 3 Python | Upstream-compatible macOS or Linux | Optional Python provider and hardware appropriate to the chosen checkpoint; a supported GPU is strongly recommended |
+| Stable Audio 3 Small, MLX | Apple Silicon macOS | Official `sa3` optimized tools and separately accepted Small SFX or Small Music weights |
+| Stable Audio 3 Small, Python | macOS or Linux | Upstream CPU path or compatible acceleration; 8 GB minimum and 16 GB system RAM suggested |
+| Stable Audio 3 Medium, Python | Linux/NVIDIA | Compatible CUDA GPU; upstream reports about 6.52 GB peak VRAM, while GERM suggests 24 GB system RAM |
 | Stability API | Any service-supported system | Network access, an operator-owned account, API key, and acceptance of service terms |
 | Native shell | macOS 13+ | Swift 5.9/Xcode command-line tools; the Python service still performs all audio work |
 
@@ -62,6 +70,7 @@ promising one universal hardware minimum.
 | --- | --- |
 | How does it work? | [Micro/Matter architecture](docs/germ_micro_architecture.md) and [provider design](docs/provider_design.md) |
 | How does it connect? | [The Listening Stack](https://sonicfield.org/stack) and [Oída integration](docs/oida-integration.md) |
+| How do I install a tested model? | [Local model setup](docs/local_setup.md) |
 | Which models and licenses apply? | [Models and licensing](docs/models-and-licensing.md) |
 | What is unfinished? | [Known limitations](#known-limitations) and [roadmap](ROADMAP.md) |
 | How can I help? | [Contribution guide](CONTRIBUTING.md) |
@@ -132,6 +141,18 @@ provenance, retention, and lineage between the organs.
 
 ## Install
 
+For a guided GERM or complete Listening Stack installation, including model
+choice, storage and memory guidance, gated-access checks, downloads, and local
+gateway configuration:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/sonicfieldlabs/listening-stack/main/install.sh | bash
+```
+
+Choose **GERM only** or **Oída + GERM** in the terminal assistant. GERM remains
+in this repository; the installer only coordinates its source, dependencies,
+provider, models, and local configuration.
+
 Requirements: Python 3.10+ and `uv`.
 
 ```bash

docs/local_setup.md

@@ -75,7 +75,21 @@ in or have not accepted the Stability AI model terms.
    uv run hf auth login
    ```
 
-3. Start the server and verify access:
+3. Download the exact checkpoint. Stable Audio's loader also downloads on first
+   use, but an explicit download makes storage and access failures visible before
+   a generation job:
+
+   ```bash
+   uv run hf download stabilityai/stable-audio-3-small-sfx
+   uv run hf download stabilityai/stable-audio-3-small-music
+   uv run hf download stabilityai/stable-audio-3-medium
+   ```
+
+   Download only the models you intend to use. The Hugging Face API currently
+   reports approximately 3.49 GB for each Small post-trained repository and
+   10.45 GB for Medium.
+
+4. Start the server and verify access:
 
    ```bash
    ./launch_germ.command
@@ -84,6 +98,23 @@ in or have not accepted the Stability AI model terms.
 
 You can run the same check from the dashboard Status section with `HF Check`.
 
+The exact model page controls. At the time of this public-alpha release, the
+three pages are gated and identify the Stability AI Community License plus
+additional component terms. GERM cannot accept those terms for an operator.
+
+### Guided Listening Stack Install
+
+The separate installer can prepare GERM or both main applications, show the
+combined disk and RAM plan, authenticate Hugging Face, download only selected
+models, and leave the gateways ready to start:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/sonicfieldlabs/listening-stack/main/install.sh | bash
+```
+
+It keeps model caches, state, logs, outputs, and local paths outside this Git
+repository.
+
 ## Apple Silicon MLX Provider
 
 ```bash

docs/models-and-licensing.md

@@ -17,11 +17,26 @@ The optional Python provider references the official
 Its repository [code license](https://github.com/Stability-AI/stable-audio-3/blob/main/LICENSE)
 is MIT at the time of this release.
 
-Stable Audio 3 model weights are separately licensed. For example, the
-[Stable Audio 3 Small SFX model card](https://huggingface.co/stabilityai/stable-audio-3-small-sfx)
-identifies the Stability AI Community License and may identify additional
-component terms. Some checkpoints are gated. The upstream model card and terms
-for the exact checkpoint always control.
+GERM's first developed local generation path uses three post-trained
+checkpoints:
+
+- [Stable Audio 3 Small SFX](https://huggingface.co/stabilityai/stable-audio-3-small-sfx)
+- [Stable Audio 3 Small Music](https://huggingface.co/stabilityai/stable-audio-3-small-music)
+- [Stable Audio 3 Medium](https://huggingface.co/stabilityai/stable-audio-3-medium)
+
+Stable Audio 3 model weights are separately licensed. These model cards are
+gated and currently identify the Stability AI Community License plus
+additional component terms. The exact page and terms for the downloaded
+checkpoint always control. Locally runnable or downloadable weights should not
+be described as unrestricted open-source software solely because the code
+repository is MIT.
+
+The upstream release supplies text-to-audio, audio-to-audio editing,
+inpainting and continuation, variable-length generation, and LoRA
+personalization. GERM places those primitives inside its graph, Micro/Matter
+modules, job system, source handling, comparison, metadata, and sonic-lineage
+workflows. The wider Listening Stack remains model-agnostic even though this
+first GERM provider is intentionally designed around Stable Audio 3.
 
 GERM does not redistribute weights or accept upstream terms for the operator.
 Before installing a model:
@@ -34,6 +49,11 @@ Before installing a model:
 The MLX installer prepares the official provider path on Apple Silicon; it
 does not change the model's license.
 
+See [Local setup](local_setup.md) for explicit `hf download` commands and the
+[Listening Stack installer](https://github.com/sonicfieldlabs/listening-stack)
+for the guided route. Neither path redistributes weights or records a Hugging
+Face token in Git.
+
 ## Stability API and Other Providers
 
 The Stability API route is opt-in and uses an operator-supplied key. Service