# DreamStack — Vision & Use Cases

> **Core insight**: DreamStack turns UI into a streamable binary protocol.
> Any `.ds` app becomes multiplayer with one line. No networking code. No database sync.
> Just `stream main on url { mode: signal }`.

---

## The Paradigm Shift

Today, a UI is a **program** running on your device. With DreamStack's bitstream,
a UI is a **signal stream** — writable by any source, receivable by any surface,
composable, replayable, and physics-native.

| Traditional | DreamStack |
|---|---|
| App runs on one device | Signal stream renders on N surfaces |
| Multiplayer = WebSocket + state sync + conflict resolution | Multiplayer = one `stream` declaration |
| Screen sharing = compressed pixels | State sharing = signal diffs (10-100x smaller) |
| Session replay = specialized SDK | Session replay = just record the bitstream |
| Physics = external library | Physics = built-in WASM engine, state streamed |

---

## The Five Primitives

DreamStack's full vision combines five independent primitives into something
that doesn't exist yet:

| Primitive | Role |
|---|---|
| **DreamStack bitstream** | UI as a streamable signal protocol |
| **Iroh P2P** | Zero-infrastructure device mesh |
| **Local LLM** | On-device intelligence, fully private |
| **Solana NFT** | Ownership, identity, payments |
| **PgFlex** | Schemaless, real-time, zero-config database |

Each is powerful alone. Combined, they produce **a portable, tradeable,
self-sovereign software company — minted as a single NFT**.

---

## Part 1 — Fundamental UI Redesigns

### 1. Ambient UI — App Without a Device

A `.ds` app runs once (server, edge, or one browser). Its signal stream renders
on every available surface simultaneously:

- **Phone** shows controls (buttons, sliders)
- **Laptop** shows detail view (data, forms)
- **TV** shows visualization (charts, physics)
- **Watch** shows status (alerts, counts)

This isn't responsive design (same app adapting). It's **one signal graph,
many renderers**. Each surface subscribes to the signals it cares about
and renders at its own resolution and frame rate.

**Why it matters**: Apps stop being "installed things" and become ambient
services that appear on whatever screen is nearest.

#### Universal Receiver Spectrum

The simplest device that can accept a DreamStack stream needs only:
a socket client, 16-byte header parsing, and any output surface.

| Device | What it consumes | Cost |
|--------|-----------------|------|
| **ESP32 + LED** | Signal frames only (`0x30`/`0x31`) — ~20 lines MicroPython | ~$4 |
| **Raspberry Pi + e-ink** | Signal frames → text render | ~$15 |
| **Any browser tab** | All 6 frame types (zero-framework HTML receiver) | Free |
| **Phone / tablet** | All frames + touch/pointer input back | Free |
| **Smart display** | Full bidirectional + physics | ~$100 |

A signal-mode stream sending `{"brightness": 0.7}` is trivially parseable
even on a microcontroller. The source decides the *meaning*, the receiver
decides the *manifestation*.

### 2. Time-Travel Interfaces

Every signal diff is a timestamped binary frame. Record the stream →
complete session recording. Scrub to any point. Fork from any moment
into a new interactive session.

- **Bug reporting**: User reports a bug → replay their exact session,
  scrub to the problem, fork it, debug live
- **Tutorials**: Record an interactive tutorial — viewers replay it
  at their own pace, with full interactivity
- **Undo/redo**: Unlimited, branching undo — fork from any point in history
- **Analytics**: Replay user sessions to understand behavior (not heatmaps —
  actual interactive replays)

### 3. UI Composition via Signal Mixing

```
source weather_api | layout dashboard | render tv
```

Take a data signal stream from a weather API. Take a layout signal from
a design tool. Take interaction signals from a game controller. Merge them
into one renderer.

- **Dashboards**: Compose N data sources into one view without backend
- **Mashups**: Overlay a chat stream on a video stream on a data stream
- **Customization**: Users pick their own layout for someone else's data

**Why it matters**: UI becomes composable like Unix pipes.

### 4. Zero-Install Instant Apps

The receiver doesn't need the app code. It just decodes the bitstream and
renders. Connect to `ws://app.example.com/stream/main` and the full
interactive UI appears. Close the tab, it's gone.

- No download, no permissions, no App Store
- The app "lives" on the source — viewers are just windows into it
- Updates are instant — the source changes, all viewers see it immediately
- Works on any device with a bitstream receiver

### 5. Physics-Native Interaction

With `ds-physics` baked into the language, UI elements aren't positioned —
they're simulated.

- Drag a panel → it continues with momentum
- Throw a card → it bounces off other cards
- Pinch a stack → items push apart with spring forces
- Resize a container → content flows like water

The physics state streams like any other signal. Multiple users can
interact with the same physics world simultaneously.

---

## Part 2 — Iroh P2P: Zero-Infrastructure Streaming

Adding Iroh P2P to the bitstream protocol removes the relay server entirely
and opens use cases that centralized infrastructure cannot serve.

| | WebSocket Relay | WebRTC | **Iroh P2P** |
|---|---|---|---|
| Discovery | Needs known URL | Needs signaling server | **Content-addressed (hash)** |
| NAT traversal | N/A (server) | ICE/STUN/TURN | **Built-in holepunching** |
| Topology | Star (relay center) | Peer pairs | **Mesh / swarm** |
| Offline-first | ❌ | ❌ | **✅ (local-first sync)** |
| Identity | URL-based | Session-based | **Cryptographic (public key)** |
| Persistence | Relay must stay up | Connection must stay open | **Content persists in network** |

### Key Use Cases

**Zero-Infrastructure Streaming** — A teacher in a classroom with no internet
runs a live `.ds` lesson to 30 student tablets over LAN via mDNS discovery.
No server. No URL. No relay cost.

**Persistent Bitstream Archives** — A recorded bitstream session becomes a
content hash. Anyone with the hash fetches and replays it from the swarm.
Bug reports, tutorials, session replays — all just 32-byte hashes. No replay server.

**Mesh Multiplayer** — N peers form a gossip mesh instead of routing through a
star topology. If one person drops, the others keep going. Latency drops
(direct paths instead of relay bounce).

**Offline-First + Sync-on-Reconnect** — Device goes offline, accumulates signal
diffs locally, reconnects → Iroh syncs missing state automatically. Field workers,
IoT sensors, home displays — all work disconnected and reconcile on reconnect.

**Cryptographic Identity** — Every Iroh node has a public key. Zero-config auth:
only whitelisted keys can subscribe to your stream. No API keys, no OAuth.

---

## Part 3 — Local LLM: Intelligent Ambient UI

Combining DreamStack + Iroh + local LLM inference produces intelligent, private,
serverless computing that renders on any surface.

### The AI That Lives in the Room

A single device (NUC, Mac Mini, laptop) runs a local LLM and joins the Iroh mesh.
Every surface in the house receives its signal stream.

- Speak to **any screen** → voice input through the mesh → local LLM processes →
  response streams as signal diffs to **every surface**
- Kitchen display shows the recipe. Phone shows the shopping list. TV shows the
  tutorial video. **One inference, N renderers.**
- All private. Nothing leaves the network. Ever.

### Distributed Inference Swarm

Multiple devices in the mesh each run small models. The signal protocol
coordinates them:

- Phone runs a fast model for intent detection
- Desktop runs a large model for generation
- No orchestration server. Models discover each other via Iroh.
- **LLM inference becomes a mesh protocol.**

### Self-Organizing Smart Spaces

The LLM observes signal streams flowing through the mesh and adapts the UI:

- Sees calendar signals → "Meeting in 10 minutes" → pushes notification to watch
- Sees IoT sensor signals → generates natural-language summary → nearest display
- Learns patterns → proactively streams control signals

No IFTTT. No Home Assistant rules. The LLM **is** the automation engine,
DreamStack **is** the I/O bus.

### Private AI Tutor

Teacher runs a `.ds` physics simulation. Students interact via tablets (touch inputs
stream back through Iroh mesh). A local LLM watches each student's interaction
stream and generates personalized hints. The hints stream as signal diffs only to
that student's device.

**Adaptive tutoring. Zero cloud. Works offline in rural schools.**

### Conversational Database (with PgFlex)

The local LLM watches PgFlex's SSE change stream and builds understanding:

- "Top 3 products last week?" → LLM generates PostgREST query → PgFlex returns → DreamStack renders
- "Alert me when inventory < 50" → LLM creates computed field + threshold → SSE fires when triggered
- PgFlex's schemaless nature means the LLM can create new fields on the fly
- **The LLM becomes a database architect that responds to natural language**

| Without LLM | With Local LLM |
|---|---|
| Streams carry **data** | Streams carry **meaning** |
| Surfaces **render** signals | Surfaces **understand** signals |
| Users **configure** automations | The system **infers** intent |

---

## Part 4 — Content-Addressed UI Distribution

This is where all primitives converge into the most radical departure from
traditional software distribution.

### Two Layers of Hash

| Layer | What | Analogy |
|---|---|---|
| **App hash** | Compiled `.ds` code (static artifact) | Torrent of an executable |
| **Stream topic** | Live signal state (running instance) | Joining a live game server |

When you share `ds:bafk2bza...`, the receiver:
1. **Fetches the app code** from the swarm (content-addressed, cached, verified)
2. **Joins the live stream** from the source (signal diffs, real-time state)

You're not distributing a file. **You're distributing a running application.**

### Why It Matters

**Link rot / platform risk** — A content-addressed app cannot go offline unless
every peer disappears. No Vercel dependency. No AWS bill surprise. No domain expiration.

**Distribution without gatekeepers** — App Store takes 30%. A `.ds` app distributed
via hash has zero distribution cost and zero gatekeepers. The app doesn't need to be
"installed" — the receiver just decodes the bitstream.

**Instant, zero-trust sharing** — Today: URL → DNS → load JS → hydrate → render
(2-5 seconds). With hash: peer already has the bundle cached → stream connects →
UI appears (~100ms).

**Versioning is automatic** — v1.0 = hash `abc123`, v1.1 = hash `def456`. Both
exist simultaneously in the swarm. No deployment. No rollback scripts. Users can
pin a version. Developers publish updates without breaking existing users.

---

## Part 5 — Solana NFT: App Ownership On-Chain

A compiled `.ds` app is small — signal-mode apps are a reactive state graph plus
view declarations. Compiled, that's kilobytes. Small enough to store on-chain.

### The NFT Structure

```
┌─────────────────────────────────────────┐
│  Solana NFT (Metaplex Core)             │
│                                         │
│  metadata: { name, creator, version }   │
│  appData:  <compiled .ds bytecode>      │
│  stream:   <iroh topic hash>            │
│  dbConfig: <pgflex connection, encrypted│
│                                         │
│  owner: 7xKXt...  (your wallet)        │
└─────────────────────────────────────────┘
```

The NFT contains:
1. **The compiled app** — `.ds` bytecode in the AppData field
2. **The stream topic** — Iroh topic hash for the live instance
3. **The database config** — PgFlex connection (encrypted, owner-only)
4. **Ownership** — enforced by Solana consensus

### What Ownership Means

Today, owning an NFT means owning a JPEG pointer. Owning a DreamStack NFT means
owning a **running application**:

**The owner controls the source stream** — Only the wallet owner can publish signal
diffs to the stream topic. Transfer the NFT → the new owner becomes the source.
The app literally changes hands like a physical object.

**The NFT IS the deployment** — No Vercel. No domain. No hosting account.
`ds run sol:ADDRESS` → fetch bytecode from chain → boot the app → join the stream.
The app's existence is guaranteed by Solana's uptime, not yours.

**Versioning is on-chain** — Update = new transaction → AppData updates.
Full version history in the transaction log. Roll back = point to a previous
transaction's data.

### Apps as Tradeable Assets

A developer builds a `.ds` app. Mints it as an NFT. Sells it. The buyer owns and
operates the app. The developer gets royalties on secondary sales (Metaplex royalties).
Build once, earn forever.

### Access-Gated Streaming

The stream topic is public, but the source checks NFT ownership:

```
stream main on iroh {
  gate: nft("NFT_ADDRESS")  -- only NFT holders can connect
}
```

- Mint 100 "access pass" NFTs for a premium dashboard
- Each holder connects to the live stream
- Trade the access pass on the open market when done
- **Subscription replaced by asset ownership**

### Composable App Economy

Because `.ds` apps are signal graphs, they compose:

```
let data = stream from nft("NFT_A")   -- a data feed
let viz  = stream from nft("NFT_B")   -- a visualization
-- Compose into a new app → mint as NFT_C
```

NFT_C depends on NFT_A and NFT_B. On-chain, this dependency is recorded.
When NFT_C earns revenue, royalties flow upstream automatically.
**An app supply chain with automatic revenue sharing, enforced by the blockchain.**

### The App Store Without Apple

| Traditional | DreamStack + Solana |
|---|---|
| Developer → App Store (30% cut) → User | Developer → Mint NFT (0.1 SOL) → User |
| Apple controls listing | Apps are permissionless |
| User pays subscription | User buys NFT (resellable) |
| App dies when company dies | App lives on-chain forever |
| No secondary market | Apps trade like assets |

---

## Part 6 — PgFlex: The Memory Layer

PgFlex solves the problem the other four primitives don't: **structured, queryable,
persistent state**. Bitstream signals are ephemeral. Iroh stores blobs but isn't a
database. The LLM reasons but doesn't remember. The NFT stores code, not data.

PgFlex is the **memory** of the system.

### Full-Stack Architecture

```
┌─────────────────────────────────────────────────────────┐
│  Solana NFT                                             │
│  • bytecode  • iroh topic  • db config                  │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│  Device (runs the app)                                  │
│                                                         │
│  DreamStack Runtime                                     │
│  ├── Signal graph (reactive state)                      │
│  ├── Local LLM (inference)                              │
│  ├── Iroh P2P (stream to N surfaces)                    │
│  └── PgFlex client (persistent storage)                 │
│       │                                                 │
│  PgFlex Stack                                           │
│  ├── PostgreSQL 17 (data lives here)                    │
│  ├── PostgREST (zero-code REST API)                     │
│  ├── SSE Proxy (real-time change events)                │
│  └── ElectricSQL (edge sync / offline)                  │
└─────────────────────────────────────────────────────────┘
```

### Self-Contained SaaS on an NFT

The NFT contains the entire stack config: app bytecode (DreamStack), stream topic
(Iroh), database schema (PgFlex — auto-discovered), LLM model reference.

Buy the NFT → `ds run sol:ADDRESS` → PgFlex boots with zero schema → the app writes
data → PgFlex auto-discovers fields → **fully functional SaaS with persistent data,
from a single NFT**.

You're not selling an app. You're selling a running business.

### Data Sovereignty by Default

PgFlex runs on YOUR device/server. The NFT says "this app stores data in PgFlex" but
doesn't dictate where. The buyer deploys their own PgFlex (`docker compose up`).
Their data never touches the developer's infrastructure.

- Buy a CRM-as-NFT → deploy PgFlex locally → all customer data stays on your server
- The app creator has zero access to your data
- **GDPR/HIPAA compliance by architecture, not by policy**

### PgFlex SSE → DreamStack Signal Bridge

PgFlex emits real-time change events via SSE (`pg_notify`). Bridge those directly
into DreamStack's signal graph:

```
let orders = pgflex.subscribe("orders")

view dashboard =
  column [
    text "New orders: {orders.length}"
    each orders as order ->
      text "{order.customer}: ${order.total}"
  ]

stream dashboard on iroh { topic: "dashboard" }
```

Database changes → PgFlex SSE → DreamStack signals → Iroh mesh → every screen
updates. **No API layer. No polling. No frontend state management.**

### Why Not Automerge / CRDTs?

CRDTs (like Automerge) solve concurrent uncoordinated writes with no central
authority. PgFlex doesn't need them because **no layer in this architecture
produces that problem**:

```
Layer               Sync mechanism          Conflicts?
────────────────────────────────────────────────────────
DreamStack signals  Version map + diffs     No — single source per stream
PgFlex writes       PostgreSQL transactions No — single DB is authority
PgFlex reads        ElectricSQL shapes      No — read-only replication
Iroh P2P            Content-addressed blobs No — immutable by hash
```

PgFlex's natural topology is **single-writer**: one PostgreSQL instance per
deployment (home, school, hospital). Within that deployment, PostgreSQL handles
write ordering via transactions. ElectricSQL handles read replication to edge
devices. SSE pushes change notifications. No write contention occurs.

For offline scenarios (two peers editing state while disconnected), the conflict
resolution happens **in the signal layer, not the database layer**:

```
Peer A ──signals──→ DreamStack merge ←──signals── Peer B
                         │
                   resolved state
                         │
                      PgFlex
                  (just persists)
```

DreamStack's signal protocol reconciles on reconnect (last-write-wins per signal,
using version maps). PgFlex materializes whatever the signal graph settles on.
Adding Automerge would be a redundant conflict resolution layer between two
systems that already don't produce conflicts.

> [!NOTE]
> If PgFlex-to-PgFlex replication across sites is ever needed, the answer is
> PostgreSQL logical replication + custom conflict handlers — not Automerge.
> Stay in the PostgreSQL ecosystem.

---

## Revenue-Generating Applications

### Tier 1: High Conviction — Clear Buyer, Clear Budget

#### A. Private AI Appliance for Regulated Industries

**Buyer**: Hospitals, law firms, financial advisors, government agencies
**Budget they already spend**: $50K-500K/year on HIPAA/SOC2-compliant cloud AI

**The product**: A box (Jetson/NUC) running local LLM + DreamStack + PgFlex. Plug into
the network. Every screen in the facility gets AI — medical record summaries on the
doctor's tablet, appointment prep on the wall display, billing codes on admin screen.
Zero data leaves the building.

**Why they pay**: Compliance. They *can't* use ChatGPT for patient data. Local LLM
solves this. DreamStack solves multi-surface delivery. PgFlex stores patient data locally.

**Revenue**: Hardware ($500-2000) + annual software license ($2K-10K/yr)
**Market**: Healthcare IT alone is $400B. Even 0.001% = $4M.
**Moat**: Multi-surface streaming. Competitors sell "local LLM in a box" but output
is one terminal. DreamStack makes it ambient.

#### B. Live Education Platform

**Buyer**: School districts, edtech resellers, corporate training departments
**Budget they already spend**: $5-15/student/year on Nearpod, Pear Deck, Kahoot

**The product**: Teacher runs interactive simulations on their laptop. Students connect
from any device — Chromebook, iPad, phone. Over LAN via Iroh when internet is down.
Local LLM provides per-student adaptive hints.

**Why they pay**: Current tools are glorified PowerPoint with polls. DreamStack
delivers actual interactive simulations + works offline (huge for rural schools).

**Revenue**: $8/student/year (undercut Nearpod at $15)
**Market**: 50M US K-12 students × $8 = $400M TAM
**Moat**: Physics-native simulations over bitstream. Offline-via-Iroh is a killer
feature for school procurement.

### Tier 2: Strong Signal — Needs Validation

#### C. Full-Stack NFT Marketplace

**Buyer**: Solo developers, indie SaaS builders, agencies
**Budget**: Currently $0 (new market) or $20-100/mo on hosting

**The product**: `ds publish --solana` mints your app as a tradeable NFT. Buyers get a
complete, deployable, intelligent app with persistent storage. Developers earn royalties
on secondary sales.

| What you sell | What the buyer gets | Price range |
|---|---|---|
| CRM NFT | App + PgFlex schema + LLM customer insights | 10-50 SOL |
| Analytics Dashboard NFT | Real-time dashboard + data layer + anomaly detection | 5-20 SOL |
| Inventory Tracker NFT | Multi-surface inventory + auto-reorder via LLM | 20-100 SOL |
| Custom App (commissioned) | Bespoke `.ds` app minted for one client | 100+ SOL |

**Revenue**: Transaction fees (2.5% of NFT sales) + pinning service
**Moat**: The NFT isn't a JPEG — it's a running application. Utility value, not speculative.

#### D. DreamStack Registry (ds.run)

**The product**: `ds publish` → app gets a content hash. Anyone with the hash
fetches and runs it from the swarm.

| Tier | What | Price |
|---|---|---|
| **Free** | Publish to swarm (ephemeral, lives as long as peers have it) | $0 |
| **Pin** | Permanent peers keep your app alive 24/7 | $5/app/month |
| **Pro** | Pin + custom domain + analytics + `ds.run` web gateway | $20/month |
| **Enterprise** | Private registry, SSO, audit log, SLA | $200/month |

The web gateway `ds.run/hash` opens the app in a browser via the WASM receiver.
Lets you share apps with people who don't have DreamStack yet.

**Moat**: The hash isn't just code — it includes the live stream topic. Pinned
apps are running, interactive, multiplayer applications. IPFS = static files.
Vercel = static sites. DreamStack Registry = live interactive apps as hashes.

#### E. Relay-as-a-Service

**Buyer**: SaaS developers building collaborative features
**Budget**: $0.10-1.00/concurrent connection/month on Liveblocks

**The product**: `npm install dreamstack` → add one `stream` declaration → instant
multiplayer. Hosted relay + Iroh P2P fallback.

**Revenue**: Usage-based, $0.05/connection/month
**Moat**: Signal-graph-aware sync (not generic pub/sub), binary efficiency,
P2P fallback = lower infrastructure cost

#### F. IoT / Operations Dashboard

**Buyer**: Factory floor managers, DevOps teams, operations centers
**Budget**: $15-50/user/month on Grafana Cloud, Datadog

**The product**: Sensor data streams via Iroh mesh to any screen. Local LLM adds
anomaly detection + natural language alerts. No cloud egress fees.

**Revenue**: $20/source/month
**Moat**: Binary efficiency + P2P = massive cost savings at scale

---

## Go-to-Market Sequence

```
Phase 1 (NOW → 3 months):   Open-source the protocol + relay.
                              Ship the "Collaborative Step Sequencer" demo.
                              → Developer attention + GitHub stars.

Phase 2 (3-6 months):        Relay-as-a-Service (E).
                              → First revenue, proves developer demand.
                              → Low cost to operate (already built).

Phase 3 (6-12 months):       Education Platform (B) + ds.run Registry (D).
                              → Pick one school district, pilot it.
                              → Add local LLM hints as differentiator.
                              → Registry grows alongside open-source adoption.

Phase 4 (12-18 months):      Private AI Appliance (A) + NFT Marketplace (C).
                              → Partner with healthcare IT reseller.
                              → Hardware margin + recurring license.
                              → NFT marketplace leverages existing Solana ecosystem.
```

**Through-line**: DreamStack's moat is **multi-surface + binary efficiency + P2P**.
Every product must lean into all three. If a product only needs one, someone else
will eat your lunch.

---

## Demo Prioritization

| Demo | Effort | Wow Factor | Revenue Signal |
|------|--------|------------|----------------|
| Collaborative Step Sequencer | 1 day | ⭐⭐⭐⭐⭐ | Relay-as-a-Service |
| Two-Tab Synced Counter | 1 hour | ⭐⭐⭐ | — |
| Physics Playground | 2 days | ⭐⭐⭐⭐ | Education |
| Collaborative Whiteboard | 2 days | ⭐⭐⭐⭐ | Design Tool |
| ESP32 Signal Receiver | 1 day | ⭐⭐⭐⭐ | IoT / Ambient UI |
| Session Replay Viewer | 3 days | ⭐⭐⭐ | Analytics |
| NFT-Minted App + ds.run | 3 days | ⭐⭐⭐⭐⭐ | NFT Marketplace |
| Multi-Surface Dashboard | 1 week | ⭐⭐⭐⭐⭐ | IoT/Operations |

**Recommended first demo**: Collaborative Step Sequencer.
It's visual, temporal, interactive, and multiplayer — proving the entire
stack in the most visceral way possible. Two browser tabs, one beat grid,
instant sync, zero networking code.

**Killer demo for NFT vision**: Publish a step sequencer as a content hash.
Share it in a tweet. Anyone who clicks the `ds.run/hash` link instantly joins
the live session — no account, no install, no server. 50 people making music
together from a 32-byte hash.

---

## Technical Foundation (Complete)

All infrastructure is built and tested:

- ✅ **Bitstream protocol**: 16-byte binary header, 6 frame types
- ✅ **Relay v1.0.0**: Multi-source routing, keyframe cache, channel GC,
  max receivers, source reconnection, 54 tests
- ✅ **WASM codec**: Shared Rust↔browser encoding (18KB)
- ✅ **WebRTC transport**: Peer-to-peer data channels with auto-fallback
- ✅ **Compiler integration**: `stream` keyword, `stream from` expression,
  `transport: webrtc` option
- ✅ **Receiver protocol**: All 6 frame types, RLE decode, exponential backoff
- ✅ **PgFlex**: Schemaless PostgreSQL, PostgREST, SSE, ElectricSQL — 110 tests
- ✅ **110+ tests, 0 failures**

### Primitives Still Needed

- ⬜ **Iroh P2P transport**: Replace WebSocket relay with Iroh mesh
- ⬜ **`ds publish`**: Mint `.ds` app as Solana NFT
- ⬜ **`ds run sol:ADDRESS`**: Fetch + boot from on-chain bytecode
- ⬜ **ds.run gateway**: Web gateway for content-addressed apps
- ⬜ **NFT-gated stream auth**: Check wallet ownership before allowing connection
- ⬜ **PgFlex signal bridge**: SSE → DreamStack signal graph
- ⬜ **Local LLM integration**: Model loading + signal-based I/O

---

## Part 7 — Impossible Use Cases (Enabled by 44KB Output)

> The JS output optimizations (DOM helpers, tree-shaking, minification) reduced
> compiled DreamStack apps to **44KB** (~8KB gzipped). At this size, an interactive
> application crosses a threshold: **it stops being infrastructure and becomes data.**
> Data can travel through streams, live on-chain, or be generated on the fly.

### 1. Frame-Zero Self-Propagating Apps

**The insight**: The app is small enough to be sent through its own data stream.

```
┌──────────────┐         ┌──────────────┐         ┌──────────────┐
│   Source      │         │    Relay     │         │   Browser    │
│  (Pi, laptop) │         │  (any relay) │         │  (viewer)    │
└──────┬───────┘         └──────┬───────┘         └──────┬───────┘
       │  FRAME 0: full HTML    │                        │
       │  (44KB compiled app)   │                        │
       │───────────────────────>│  caches it             │
       │  FRAME 1: {temp: 72.4} │                        │
       │───────────────────────>│                        │
       │                        │   Someone opens URL    │
       │                        │<──────── GET /s/a8f3c  │
       │                        │  serves FRAME 0        │
       │                        │───────────────────────>│ renders app
       │                        │   WS auto-connects     │
       │                        │<──────── upgrade ──────│
       │  FRAME N: {temp: 73.1} │                        │
       │───────────────────────>│───────────────────────>│ live update
```

One command: `dreamstack stream sensors.ds --bootstrap`
One URL: `https://relay.dreamstack.dev/s/a8f3c`
No CDN, no hosting, no separate frontend deployment.

**Why impossible today**: React (200KB+) is too large to push through a WebSocket.
Firebase/Supabase require pre-deployed frontends. No existing framework can travel
through its own data channel.

### 2. Solana Accounts as Reactive Signal Sources

**The insight**: DreamStack signals map directly to on-chain state.
The wallet balance IS a reactive signal. A transfer IS a signal mutation.
**The blockchain replaces the relay server.**

This is NOT "store HTML on-chain" (anyone can do that). The signals themselves
are sourced from on-chain state. The app reads its state FROM the blockchain
and writes state BACK via transactions.

```
User clicks "Send 0.5 SOL"
       │
       ├─→ balance signal reacts
       ├─→ buildTransaction(transfer(recipient, 0.5 SOL))
       ├─→ wallet.signAndSend(tx)    ← Phantom popup
       │
       ▼
   Solana validators confirm (~400ms)
       │
       ├─→ accountSubscribe fires: balance changed
       ▼
   Sender's UI: balance spring-animates down
   Recipient's UI: balance spring-animates up
   Both update with zero infrastructure between them
```

**Zero infrastructure.** No relay, no WebSocket server, no database.
The blockchain IS the pub/sub layer, the state store, and the audit log.

| What you'd need  | React + Firebase | **DreamStack + Solana** |
|---|---|---|
| Frontend hosting  | Vercel ($20/mo)  | **On-chain ($0.12 once)** |
| Real-time sync    | Firebase ($25/mo)| **accountSubscribe (free)** |
| Backend server    | Cloud Functions  | **None** |
| Audit log         | Build it yourself| **Solana Explorer** |
| Total infra       | 3-4 services     | **0 services** |

**Why impossible today**: No frontend framework has a signal system that maps
to Solana account bytes. React/Vue call the chain — the chain doesn't push state
back reactively. DreamStack closes the loop: signals ↔ account data ↔ subscribers.

### 3. AI-Generated Live Applications

**The insight**: An LLM generates 20 lines of `.ds` source → compiles in <100ms
→ produces 44KB HTML → pushes to relay as frame 0. The AI responds with a
**running application** instead of text.

```
User:  "Show me a live dashboard of my Solana validator"

AI:    [generates 30 lines of .ds — 500 bytes]
       [compiles to 44KB HTML in 80ms]
       [pushes to relay as frame 0]

       → Here's your dashboard: relay.dreamstack.dev/s/x9f2
         (already live, already streaming real data)
```

The AI doesn't link you to a website. The AI doesn't generate code for you to
deploy. **The AI's response IS a live, streaming, interactive application.**

**Why impossible today**: ChatGPT produces static text. Code Interpreter produces
screenshots. V0/Bolt generate code you deploy yourself. No AI can respond with a
live streaming app because existing frameworks require build infrastructure and
are too large for inline delivery.

### The Unifying Principle

> When an application is small enough, it stops being infrastructure and becomes
> data. Data can travel through streams (frame-zero), live on-chain (Solana UI),
> or be generated on the fly (AI apps). DreamStack apps crossed that threshold —
> they're small enough to go anywhere data goes.