Architecture Overview

GEIANT is a four-layer stack. Each layer is independently useful, but the power comes from their composition — identity is grounded in geography, jurisdiction is enforced by delegation, and every operation is cryptographically chained.

The Four Layers

┌─────────────────────────────────────────────┐
│  Layer 3 — Perception Runtime               │
│  Satellite tiles, weather, embeddings       │
│  MCP-native tools, GPU inference            │
├─────────────────────────────────────────────┤
│  Layer 2 — Agent Identity Protocol          │
│  Ed25519 identity, delegation certificates  │
│  Breadcrumb chains, trust tiers, compliance │
├─────────────────────────────────────────────┤
│  Layer 1 — GEP (Geographic Addressing)      │
│  569 trillion addressable points            │
│  Open infrastructure, no token gate         │
├─────────────────────────────────────────────┤
│  Layer 0 — H3 Hexagonal Grid               │
│  Uber's geospatial indexing system          │
│  Resolution 0–15, hierarchical              │
└─────────────────────────────────────────────┘

Layer 0 — H3 Grid

The foundation is Uber's H3 hexagonal hierarchical spatial index. H3 divides the Earth's surface into hexagonal cells at 16 resolutions (0–15). GEIANT uses:

• Resolution 5 (~253 km²) for jurisdictional boundaries in delegation certificates
• Resolution 7 (~5.16 km²) for agent territorial binding
• Resolution 10 (~0.015 km²) for precise operation logging

H3 cells are deterministic, compact (64-bit integer), and support parent/child/neighbor traversal — ideal for jurisdiction checks that need to run in microseconds.

Layer 1 — GEP (GeoEpoch Protocol)

GEP provides geographic addressing at planetary scale. At resolution 15, GEP offers approximately 569 trillion unique addressable points — more than enough to assign an address to every square meter of Earth's surface.

GEP is open infrastructure with no token monetization at this layer. It serves the same role as IP addressing: a shared coordinate system that everything above depends on.

Layer 2 — Agent Identity Protocol

This is where GEIANT's core value lives. The Agent Identity Protocol provides:

Ed25519 Identity — Every agent gets a cryptographic keypair. The public key is the identity. No registry, no username, no OAuth token. The same key signs breadcrumbs, delegation certificates, and Stellar transactions.

Delegation Certificates — A human principal signs a certificate authorizing an agent to operate within specific H3 cells, using specific capabilities (facets), for a specific time period. Certificates are verified offline — no server round-trip needed.

Virtual Breadcrumbs — Every tool call produces a signed, SHA-256-chained breadcrumb recording what happened, where, when, and under whose authority. The chain is append-only and tamper-evident.

Trust Tiers — Agents progress through five tiers (Provisioned → Observed → Trusted → Certified → Sovereign) based on operational history. Each tier unlocks additional capabilities.

Epoch Rollups — Breadcrumbs are periodically Merkle-rolled into epochs — compact summaries with a Merkle root over all block hashes in the period. This enables efficient long-term auditing.

Layer 3 — Perception Runtime

The perception layer provides MCP-native tools for geospatial AI:

• perception_fetch_tile — Fetch Sentinel-2 L2A satellite tiles via STAC API
• perception_classify — Run EO foundation models (Prithvi-EO-2.0) for flood/land-use classification
• perception_embed — Generate Clay v1.5 geospatial embeddings (planned)
• perception_weather — Historical + forecast weather via Open-Meteo ERA5

Every perception tool call is automatically wrapped by the audit middleware — a breadcrumb is dropped with jurisdiction and delegation checks enforced before execution.

Data Flow

A typical agent operation flows through all four layers:

1. Agent receives task: "Check flood risk for Rome grid cell"
2. Layer 2: Preflight check
   ├── Is delegation certificate still valid? ✅
   ├── Is H3 cell within authorized territory? ✅
   ├── Is 'energy' facet authorized? ✅
   └── Is tool whitelisted? ✅
3. Layer 3: Execute perception_weather(851e8053fffffff)
4. Layer 2: Drop breadcrumb
   ├── SHA-256 context digest of tool input + output
   ├── Chain to previous block hash
   ├── Ed25519 sign the block
   └── Write to Supabase
5. Agent returns result to orchestrator

Cryptographic Guarantees

Property	Mechanism
Identity	Ed25519 keypair — no central registry
Authorization	Delegation certificate signed by human principal
Immutability	SHA-256 hash chain — tamper with one block, all subsequent blocks invalid
Non-repudiation	Ed25519 signature on every breadcrumb
Jurisdiction	H3 cell checked against delegation certificate pre-flight
Auditability	Merkle-rolled epochs with compliance report generation

Deployment Architecture

┌──────────────┐     SSE/MCP      ┌─────────────────────┐
│  AI Client   │ ◄──────────────► │  mcp-perception     │
│  (Claude,    │                  │  Railway (Node.js)   │
│   LangChain) │                  │  Port 8080           │
└──────────────┘                  └──────┬──────────────┘
                                         │
                    ┌────────────────────┼────────────────┐
                    │                    │                 │
              ┌─────▼──────┐     ┌──────▼──────┐   ┌─────▼──────┐
              │ Supabase   │     │ RunPod GPU  │   │ Open-Meteo │
              │ PostgreSQL │     │ RTX A40     │   │ ERA5 API   │
              │ + PostGIS  │     │ Prithvi-EO  │   │            │
              └────────────┘     └─────────────┘   └────────────┘

The mcp-perception service runs on Railway, exposes SSE for MCP clients, and writes audit data to Supabase. GPU inference runs on RunPod serverless endpoints.