Oysterworld

Technical Architecture

The Sensory Layer for Physical Intelligence — Web 4.0 Infrastructure

Version 1.0 March 2026 Classification: Public

1 System Overview

Oysterworld is building the sensory infrastructure for Physical Intelligence and the Web 4.0 era. The platform connects a distributed network of physical devices—smartphones, wearables, AI glasses, and pendants—into a unified data pipeline where every sensor reading is cryptographically verified through Physics-Informed Neural Networks (PINNs). The resulting high-fidelity, real-world data flows into an open marketplace where AI agents and applications consume verified physical-world intelligence. With over 25,000 DePIN nodes already deployed, Oysterworld transforms passive consumer hardware into active participants in a decentralized sensing economy.

25K

Nodes Deployed

Device Types

0.35

PINNs MAE

Architecture Layers

2 Architecture Diagram

Layer 5 — Consumption

AI Agent & Application Layer

Third-party AI agents and dApps consume verified real-world data

AI Agents dApps Analytics Platforms Research APIs

↓ ↑

Layer 4 — Distribution

API & Data Marketplace

Open marketplace for querying, licensing, and streaming verified data

REST / GraphQL API Data Licensing Streaming Feeds Access Control

↓ ↑

Layer 3 — Storage & Indexing

Verified Data Storage

Indexed, searchable repository of PINNs-verified sensor data

Time-Series DB Spatial Index (H3) On-Chain Proofs IPFS / Filecoin

↓ ↑

Layer 2 — Verification

PINNs Engine

Physics-Informed Neural Networks validate data against physical laws

PDE Residual Loss Anomaly Detection Cross-Sensor Correlation Confidence Scoring

↓ ↑

Layer 1 — Collection

Data Pipeline (OpenClaw)

On-device aggregation, normalization, and encrypted transmission

OpenClaw Agent Sensor Fusion Edge Compression TLS Transport

↓ ↑

Layer 0 — Physical

Device Layer (25K+ Nodes)

Distributed hardware generating raw sensor data at the edge

ClawPhones Puffy Health ClawGlasses AI Pendant Third-Party Devices

3 Layer Descriptions

3.1 — Device Layer

The physical foundation of the network. Each device operates as an autonomous DePIN node, continuously generating sensor telemetry from its embedded hardware. The heterogeneous device fleet provides broad environmental coverage across geographies and modalities.

Device	Sensors	Status	Nodes
ClawPhones	GPS, accelerometer, gyroscope, barometer, ambient light, proximity, magnetometer, camera, microphone	Deployed	25,000
Puffy Health	PPG heart rate, SpO2, skin temperature, 3-axis accelerometer, gyroscope	Pre-order	10,000 reserved
ClawGlasses	RGB camera, IMU, ambient light, proximity, bone-conduction audio	Beta 2026	—
AI Pendant	Microphone array, BLE beacon, accelerometer, temperature	Coming Soon	—
OpenClaw (Software)	All Android-accessible sensors via software agent	Available	Included in ClawPhones

3.2 — Data Pipeline (OpenClaw)

OpenClaw is the open-source on-device agent that transforms raw sensor readings into structured, transmission-ready data packets. Running natively on Android, it handles the full ingestion lifecycle from capture through delivery.

Sensor Fusion: Combines readings from multiple sensors into unified observation frames with synchronized timestamps and device attestation metadata.
Edge Normalization: Applies calibration offsets, unit standardization, and coordinate-system alignment before data leaves the device.
Adaptive Compression: Reduces payload size by 60–80% using delta encoding and quantization, preserving fidelity for PINNs verification.
Encrypted Transport: All data is transmitted over TLS 1.3 with certificate pinning. Payloads are signed with device-specific keys for provenance.
Offline Buffering: Queues data locally during connectivity gaps and synchronizes when the network is available, ensuring zero data loss.

3.3 — PINNs Verification Engine

The PINNs (Physics-Informed Neural Networks) Engine is the core trust layer. It validates incoming sensor data against known physical laws, catching spoofed readings, faulty hardware, and adversarial inputs before they enter the verified data store.

PDE Residual Loss: Neural networks are trained with physics-based loss functions encoding conservation laws (energy, momentum, thermodynamics). Data violating these constraints is flagged.
Cross-Sensor Correlation: Readings from co-located sensors are checked for physical consistency (e.g., barometric pressure must correlate with reported altitude).
Anomaly Detection: Statistical and learned models identify outliers, sensor drift, and sudden discontinuities that indicate hardware fault or data manipulation.
Confidence Scoring: Each verified data point receives a continuous confidence score (0–1) based on PINNs residual, sensor agreement, and historical reliability of the source node.
Performance: Current PINNs models achieve 0.35 MAE on benchmark verification tasks, with sub-second inference latency per observation batch.

Why PINNs? Traditional ML models learn purely from data and can be fooled by adversarial inputs that are statistically plausible but physically impossible. PINNs embed the laws of physics directly into the network architecture, making it fundamentally harder to generate fake data that passes verification.

3.4 — Storage & Indexing

Verified data is persisted in a multi-tier storage architecture optimized for both real-time queries and long-term archival. Cryptographic proofs are anchored on-chain for immutable auditability.

Time-Series Database: High-throughput ingestion of sequential sensor data with efficient range queries and downsampling for historical analysis.
Spatial Indexing (H3): All geolocated data is indexed using Uber's H3 hexagonal grid at resolution 9, enabling efficient spatial queries, coverage mapping, and geographic aggregation.
On-Chain Proof Anchoring: Merkle roots of verification batches are periodically committed to a public blockchain, providing tamper-proof audit trails without storing raw data on-chain.
Decentralized Archival: Long-term storage leverages IPFS for content-addressed redundancy, with Filecoin deals for persistence guarantees beyond the hot-storage tier.

3.5 — API & Data Marketplace

The marketplace exposes verified sensor data to external consumers through programmatic APIs and a licensing framework. AI agents, researchers, and application developers can query, stream, and license data based on type, geography, time range, and confidence level.

Query API (REST / GraphQL): Flexible endpoints for filtering data by sensor type, H3 cell, time window, and minimum confidence threshold.
Streaming Feeds: WebSocket and webhook-based real-time feeds for latency-sensitive consumers such as autonomous agents and monitoring systems.
Data Licensing: Granular licensing model allowing consumers to purchase access by volume, time period, geography, or data type. Node operators earn proportional rewards.
Access Control: API key and OAuth 2.0 authentication with role-based access control, rate limiting, and usage metering per consumer.

4 Security & Privacy

Domain	Mechanism	Details
Device Identity	Hardware attestation	Each node holds a unique cryptographic key pair provisioned at enrollment. All data payloads are signed at the device level, establishing tamper-evident provenance.
Transport Security	TLS 1.3 + certificate pinning	End-to-end encryption between device and ingestion endpoints. Certificate pinning prevents MITM attacks even on compromised networks.
Data Integrity	PINNs + Merkle proofs	Physics-based verification rejects spoofed data. Merkle roots anchored on-chain provide immutable proof of the verified dataset at any point in time.
User Privacy	Differential privacy + k-anonymity	Location data is generalized to H3 cells before storage. Personal identifiers are stripped at the edge. Aggregation queries enforce k-anonymity thresholds.
Access Control	OAuth 2.0 + RBAC	API consumers authenticate via OAuth 2.0. Role-based policies control data granularity and query scope. All access is logged and auditable.
Open Source	Public audit	Core components (OpenClaw agent, PINNs models) are open-sourced on GitHub, enabling community review and independent security auditing.

5 Scalability

The architecture is designed to scale horizontally at every layer, from device enrollment through data consumption. The following mechanisms ensure the system grows gracefully as the network expands from 25K to millions of nodes.

Horizontal Scaling by Layer

Layer	Scaling Strategy	Current Capacity
Device Layer	Permissionless enrollment. Any compatible device can join the network by installing OpenClaw and completing attestation. No central bottleneck for onboarding.	25K active nodes
Data Pipeline	Edge-first processing reduces server-side load. Ingestion endpoints scale horizontally behind load balancers. Backpressure mechanisms prevent overload.	Millions of observations/day
PINNs Engine	Verification is stateless and parallelizable. GPU-accelerated inference workers scale independently. Batch processing amortizes overhead across observations.	Sub-second per batch
Storage	Partitioned by time and H3 cell. Hot data in time-series DB with automatic tiering to cold storage. Decentralized archival (IPFS/Filecoin) offloads long-term persistence.	Petabyte-ready
API / Marketplace	Stateless API servers behind CDN and load balancer. Read replicas for query scaling. Rate limiting and caching reduce redundant computation.	Thousands of concurrent consumers

Growth Roadmap

Near-term (2026): Scale to 100K+ nodes with Puffy Health and ClawGlasses device launches. Expand geographic coverage across 50+ countries. Launch public marketplace beta for AI agent consumption.

Mid-term (2027): Target 1M+ nodes. Introduce federated PINNs verification for reduced latency. Expand device ecosystem through third-party hardware partnerships and SDK availability.