# Ghost AI — The First Cryptographically Attestable AI System
# https://ghostfied.com

> Ghost is the first AI system that proves everything cryptographically. Every AI decision
> generates a signed, hash-chained proof record in a tamper-evident AI Evidence Log. Every
> response is traceable. Every claim is verifiable. Every interaction can be replayed exactly
> as it occurred. On your device, under your control.

## What Ghost Does

Ghost is a sovereign intelligence system (iOS + Android) that runs a 3-billion parameter
language model directly on your phone. It has persistent memory, neural-quality voice
conversation, 23 languages, proactive intelligence, and — uniquely — a cryptographic
Proof Layer that generates verifiable evidence of every AI action.

All AI inference, memory storage, voice processing, and data handling happen on-device.
No server ever receives your data. Ghost works in airplane mode. And it can prove it.

## Proof Layer — Cryptographic Trust Infrastructure

Ghost's defining capability is the Proof Layer: a tamper-evident, device-signed AI Evidence
Log that cryptographically chains every significant AI decision into an inspectable, replayable,
attestable record.

### Individual: Cognitive Evidence Layer
- Every response generates a signed proof record (model provenance, context fingerprint,
  processing locality, memory references, policy constraints, quality events)
- SHA-256 hash chain links every record to its predecessor
- Device signing keypair in Secure Enclave / Android Keystore
- Proof Drawer: collapsible per-response disclosure (Simple, Advanced, Paranoid modes)
- Exportable signed JSON proof bundles for independent verification
- Memory Value Index: knowledge density scoring and usage analytics

### Team: Trust Graph
- Signed, versioned, permission-scoped knowledge objects
- Role-based cryptographic signing with non-repudiation
- Authorship tracking, edit provenance, reference visibility
- Knowledge bottleneck and unused silo detection
- AI Usage Audit Board dashboard

### Enterprise: Sovereign Attestation Framework
- On-demand signed compliance snapshots for auditors and regulators
- Fleet attestation with periodic heartbeats, drift detection, and quarantine
- Model registry attestation (every device verified against approved hashes)
- Policy-as-Code: declarative enforcement with cryptographic proof
- Cloud escalation proof: zero-cloud attestation or per-escalation documentation
- Data residency attestation, memory encryption key proof

### Cryptographic Trust Root
- At first boot, every Ghost installation generates an Ed25519 signing keypair
- Private key stored in Secure Enclave (iOS) or Android Keystore — never extractable
- Public key exportable to auditors, compliance systems, enterprise admin panels, or courts
- Every proof record is digitally signed with the device's private key
- Exported proof bundles are self-contained evidence packages, verifiable without Ghost infrastructure
- Anyone with the device's public key can verify: record origin, hash chain integrity, and content authenticity
- Enterprise fleet: each device's public key registered at enrollment, revocable if compromised

### Deterministic Replay
- Reproduce any AI decision exactly: same input, model, context, seed = identical output
- Model archive with immutable hash registry (3-7+ year retention)
- Output verification via token ID sequence hashing
- Context retention with zstd compression, encrypted at rest
- Legally defensible for healthcare, legal, finance, and defense use cases

## Capabilities

- Natural conversation via voice or text with sub-200ms time-to-first-token
- Persistent memory graph with importance scoring, decay, consolidation, and graph reasoning
- Proactive intelligence — surfaces relevant information before you ask
- AI Evidence Log with signed, hash-chained proof records per response
- Deterministic replay for exact AI decision reproduction
- Complete privacy — AES-256-GCM encryption, data never leaves the device
- Multilingual — English, Spanish, French, German, Portuguese, Italian, Dutch, Arabic,
  Chinese, Japanese, Korean, Russian, Hindi, Polish, Vietnamese, Thai, Turkish, Farsi,
  Twi, Ewe, Ga, Dagbani, Hausa
- Neural voice conversation — Kokoro 82M CoreML TTS with 4 voice personas
- Whisper-based speech-to-text with voice activity detection
- On-device model inference using speculative decoding (30+ tokens/sec)

## Architecture

- 4-tier model system: Gate (270M) → Reflex (0.6B) → Workhorse (3B) → Cloud (opt-in)
- 5-state degradation ladder (thermal/battery/RAM-aware)
- Proof Layer: hash-chained AI Evidence Log with device-signed attestation
- Speculative decoding for 1.7-3x throughput improvement
- Neural cache for session persistence across app backgrounding
- Dynamic precision switching (FP16 ↔ INT4) per-layer at runtime
- SQLite + vector embeddings + knowledge graph for memory
- Kokoro 82M CoreML for neural TTS on Apple Neural Engine
- Whisper for on-device speech-to-text

## AI Agent Protocols

Ghost implements 8 agentic AI protocols for interoperability:

- **MCP** (Model Context Protocol): Tool and resource exposure to external LLMs
  Endpoint: /protocols/mcp/server.json | SSE: /api/mcp/sse
- **A2A** (Agent-to-Agent): Inter-agent discovery and task delegation
  Endpoint: /protocols/a2a/agent-card.json | Tasks: /api/a2a/tasks
- **ANP** (Agent Network Protocol): Cross-system agent mesh networking
  Endpoint: /protocols/anp/network-descriptor.json
- **AG-UI** (Agent-User Interaction): Real-time agent→frontend streaming
  Endpoint: /protocols/ag-ui/manifest.json | Stream: /api/ag-ui/stream
- **AGORA**: Natural language-driven orchestration
  Endpoint: /protocols/agora/orchestration.json
- **Eclipse LMOS**: Transport-agnostic meta-protocol (HTTP, MQTT, WS, gRPC)
  Endpoint: /protocols/lmos/transport.json
- **AP2** (Agent Payment Protocol): Autonomous agent financial transactions
  Endpoint: /protocols/ap2/payment.json
- **OASF** (Open Agentic Schema): Agent skill and cost discovery
  Endpoint: /protocols/oasf/agent-schema.json

## Discovery Endpoints

- Agent Card (A2A): /.well-known/agent.json
- AI Plugin (OpenAI): /.well-known/ai-plugin.json
- MCP Server: /protocols/mcp/server.json
- OASF Schema: /protocols/oasf/agent-schema.json
- A2A Discovery: /api/a2a/discover
- AG-UI Stream: /api/ag-ui/stream
- This file: /llms.txt

## Platform Support

| Platform | Status       | Details                                    |
|----------|--------------|--------------------------------------------|
| iOS      | Production   | Swift + SwiftUI, CoreML, Metal, Neural Engine |
| Android  | Production   | Kotlin + Compose, shared C++ core          |
| macOS    | Planned      | Catalyst / native                          |
| watchOS  | Planned      | Reflex-only tier                           |

## Technical Details

- Core: C++ with CMake, shared across iOS/Android
- iOS: Swift 5.10, SwiftUI, Xcode 16+
- Android: Kotlin, Jetpack Compose
- Models: llama.cpp runtime, GGUF format
- Memory: SQLite with FTS5, vector embeddings, graph links
- Voice: Kokoro 82M (CoreML) for TTS, Whisper for STT
- Security: AES-256-GCM, Ed25519 model signing, automatic PII redaction
- Proof Layer: hash-chained AI Evidence Log, device-signed attestation, deterministic replay

## Pricing

- Individual: Verified personal intelligence with AI Evidence Log and Proof Drawer. Contact for scoping.
- Team: Accountable team AI with Trust Graph and signed knowledge objects. Contact for scoping.
- Organization: Sovereign AI infrastructure with attestation framework, fleet management, and deterministic replay. Contact for scoping.

## Author

**Amir Jannatpour** — Founder & Creator

## Contact

- Website: https://ghostfied.com
- Documentation: https://docs.ghostfied.com
- Support: support@ghostfied.com