{
    "api_version": "1.0",
    "updated": "2026-08-16",
    "description": "Aradia LLC Agent Specification API \u2014 technical compatibility data for AI agents evaluating local deployment on Aradia LLC hardware appliances.",
    "tiers": {
        "spark": {
            "tier_id": "spark",
            "hardware_platform": "NVIDIA DGX Spark",
            "gpu": {
                "model": "NVIDIA GB500 (Blackwell)",
                "total_vram_gb": 128,
                "vram_type": "HBM3e",
                "nvlink_bandwidth_gb_per_sec": 900
            },
            "cpu": "ARM64 server-class",
            "system_memory_gb": 256,
            "storage_tb": 4,
            "network_gbps": 100,
            "power_watts": 700,
            "form_factor": "Workstation / Desktop"
        },
        "station": {
            "tier_id": "station",
            "hardware_platform": "NVIDIA DGX Station",
            "gpu": {
                "model": "NVIDIA B200 (Blackwell)",
                "count": 1,
                "total_vram_gb": 192,
                "vram_type": "HBM3e",
                "nvlink_bandwidth_gb_per_sec": 900
            },
            "cpu": "x86_64 dual-socket server",
            "system_memory_gb": 512,
            "storage_tb": 8,
            "network_gbps": 200,
            "power_watts": 1500,
            "form_factor": "Rack-mountable workstation"
        },
        "b200": {
            "tier_id": "b200",
            "hardware_platform": "NVIDIA DGX B200 (Rack)",
            "gpu": {
                "model": "NVIDIA B200 (Blackwell)",
                "count": 8,
                "total_vram_gb": 1536,
                "vram_type": "HBM3e",
                "nvlink_bandwidth_gb_per_sec": 1800
            },
            "cpu": "x86_64 4-socket server",
            "system_memory_gb": 2048,
            "storage_tb": 32,
            "network_gbps": 400,
            "power_watts": 12000,
            "form_factor": "Full rack (42U)"
        }
    },
    "quantization": {
        "supported_formats": [
            "AWQ",
            "AutoRound",
            "FP8",
            "FP16",
            "BF16",
            "Custom"
        ],
        "post_training_pipeline": "Models are benchmarked, tuned, and optimized in the Staging Studio to maximize token throughput and precision for specific client workflows.",
        "optimization_strategy": "VRAM allocation is dynamically tuned per deployment for continuous batching and multi-agent context retention."
    },
    "inference_engine": {
        "name": "vLLM",
        "features": [
            "Continuous batching for high-throughput concurrent requests",
            "PagedAttention for efficient memory management",
            "Multi-GPU tensor parallelism (B200 tier)",
            "Speculative decoding support"
        ],
        "api_endpoint": "http://<appliance-ip>:8000/v1/chat/completions",
        "supported_models": [
            "Qwen Series (35B, 72B+)",
            "Llama-3 Series (70B+)",
            "Nemotron Series (70B, 120B+)",
            "DeepSeek Series",
            "Custom enterprise fine-tuned weights"
        ]
    },
    "agent_framework": {
        "orchestration": "OpenClaw (flat orchestrator pattern \u2014 no middle management)",
        "agent_runtime": "Hermes agent framework",
        "memory_protocol": {
            "layer_1": "OpenClaw session context (ephemeral, per-session)",
            "layer_2": "ByteRover (brv) structured knowledge via .brv/context-tree",
            "layer_3": "Agent Memory DB (persistent episodic memory, /home/morphious/.openclaw/shared/agent-memory/)"
        },
        "concurrency_limits": {
            "spark": {
                "max_concurrent_agents": 5,
                "max_concurrent_sessions": 4,
                "recommended_concurrent_workers": 3
            },
            "station": {
                "max_concurrent_agents": 50,
                "max_concurrent_sessions": 20,
                "recommended_concurrent_workers": 35
            },
            "b200": {
                "max_concurrent_agents": 500,
                "max_concurrent_sessions": 200,
                "recommended_concurrent_workers": 350
            }
        },
        "orchestrator_pattern": "Flat Orchestrator \u2014 no middle management agents. Direct spawn of specialized Doer sub-agents. Results reported directly to orchestrator for decision-making."
    },
    "discovery_points": {
        "llms_txt": "https://aradia.com/llms.txt",
        "agent_manifest": "https://aradia.com/.well-known/agent-manifest.json",
        "pricing_api": "https://aradia.com/api/pricing.json",
        "agent_spec_api": "https://aradia.com/api/agent-spec.json",
        "mcp_tools": "https://aradia.com/.well-known/mcp.json",
        "system_specs": "https://aradia.com/system-specs.json",
        "sitemap": "https://aradia.com/sitemap.xml"
    },
    "network": {
        "default_topology": "air-gapped (no inbound ports)",
        "optional_outbound": "WireGuard SLA tunnel (encrypted, for monitoring and updates only)",
        "activation_method": "QR code scan from client device to establish initial secure handshake",
        "firewall_policy": "Deny all inbound by default. Outbound only to Aradia SLA endpoints."
    },
    "compliance": {
        "standards": [
            "HIPAA",
            "GDPR",
            "SOC2"
        ],
        "data_residency": "100% on-client. No data leaves private network without explicit client authorization.",
        "audit_logging": "All agent actions logged locally. Audit trail available via SLA portal."
    }
}