Live Platform: physicsworldmodel.org · Block Explorer: explorer.physicsworldmodel.org

🪙 PWM mainnet went live on Base on 2026-05-22. 9 smart contracts deployed, Basescan-verified. Browse on-chain Principles, Specs, Benchmarks, and Certificates at explorer.physicsworldmodel.org.

Physics World Model (PWM) is an open protocol for AI4Science — providing verified Solutions, Benchmarks, Specs, and Principles for scientific domains where AI must respect physical laws.

PWM started from Computational Imaging — the richest domain for testing physics-aware AI — and is expanding to cover all domains where measurement physics governs what AI can know.

Every PWM artifact lives in a strict hierarchy:

Principles
  └── Specs          (derived from Principles)
        └── Benchmarks     (measurable tests for each Spec)
              └── Solutions / Certificates   (verified implementations)

All four layers are recorded on-chain on Base. Every Certificate is verifiable by anyone — no trust in a central authority required.

Science needs benchmarks that cannot be gamed. Current AI leaderboards:

• Are hosted on private servers (can be edited or removed)
• Measure accuracy on static datasets (models overfit over time)
• Have no cryptographic verification of results
• Are siloed by domain (no shared protocol across physics disciplines)

PWM solves all four problems:

• On-chain permanence — Principles, Specs, Benchmarks, and Certificates are stored on Base. No central party can alter or erase them.
• Prospective evaluation — benchmarks are sealed before submission; data is not public until after the round closes.
• Cryptographic certificates — every verified Solution is a hash-linked on-chain Certificate. The verification is public and reproducible.
• Domain-agnostic protocol — the same four-layer stack applies to any physics domain. Computational Imaging is the founding domain; others follow.

Computational Imaging is the launching domain for PWM because it uniquely combines:

• Forward physics that is fully specifiable (optics, MRI, CT, acoustic, electron)
• Inverse problems where AI solutions must respect physical operators
• Measurable ground truth via simulation + calibrated lab instruments
• Clinical stakes (MRI, CT, OCT) that demand verified, auditable AI

This repository is the research toolkit and algorithm catalog for the Computational Imaging domain:

The on-chain protocol (mainnet/) and this research toolkit are complementary: the toolkit defines and evaluates methods; the protocol certifies and permanently records them.

pip install -U pip
pip install -e packages/pwm_core
pip install -e packages/pwm_AI_Scientist

# Microscopy
pwm run --prompt "SIM structured illumination, 3 angles, 3 phases, live cell"

# Compressive imaging
pwm run --prompt "CASSI spectral imaging, 28 bands, coded aperture"

# Medical imaging
pwm run --prompt "CT sparse view, 90 angles, low dose"
pwm run --prompt "MRI accelerated, 4x undersampling, parallel imaging"

# View results
pwm view runs/latest

# Score a method on a modality
pwm evaluate --method my_solver --modality cassi --track correct

# Run the full 4-scenario protocol
pwm evaluate --method my_solver --modality cassi --scenarios I,II,III,IV

172 imaging modalities across 5 physical carriers:

Full catalog: docs/modality_catalog.md

PWM's algorithm toolkit is grounded in two theoretical results:

Every imaging forward model admits an ε-approximate representation as a typed DAG over exactly 11 imaging primitives: Propagate, Modulate, Project, Encode, Convolve, Accumulate, Detect, Sample, Disperse, Scatter, Attenuate.

These are the physics dialect — they describe how photons, X-rays, acoustic waves, and electrons interact with matter to produce measurements. Basis growth saturates at K=11 across 172 modalities.

Two primitive namespaces. The 11 imaging primitives are the physics-level language for forward models. A separate set of 12 general computational primitives (Differentiate, Integrate, Solve, Evaluate, Evolve, Transform, Project, Sample, Couple, Constrain, Discretize, Optimize) forms the computational substrate used to discretize, simulate, invert, and validate those forward models — see "A Judge Agent Closes the Reliability Gap in AI-Generated Scientific Simulation" (Yang, 2026), Table 3.

Every reconstruction failure decomposes into three root causes:

Gate 3 dominates across all validated modalities. Autonomous correction recovers +0.8 to +10.7 dB without retraining.

Reference: "Ten Primitives and Three Gates: The Universal Structure of Computational Imaging" (Yang & Yuan, 2026)

The on-chain layer lives in mainnet/:

• 9 smart contracts on Base mainnet, Basescan-verified
• 531 genesis Principles on Base Sepolia testnet
• 533 genesis Benchmarks on Base Sepolia testnet
• Live explorer at explorer.physicsworldmodel.org

Grant applications and funding strategy: grants/

Detailed roadmap at physicsworldmodel.org/roadmap.

Four ways to contribute:

See CONTRIBUTING.md for details.

Physics_World_Model/
├── packages/           — pwm_core, pwm_AI_Scientist Python packages
├── mainnet/            — on-chain protocol: contracts, addresses, deploy logs
├── grants/             — grant applications and funding coordination
├── benchmarks/         — benchmark definitions and results
├── datasets/           — dataset adapters and registry
├── rails/              — SolveEverything gear implementations
├── docs/               — protocol and modality documentation
├── examples/           — notebooks and quickstart
├── contrib/            — solver registry, community contributions
└── papers/             — preprints and references

MIT — see LICENSE

@software{yang2026pwm,
  author  = {Yang, David},
  title   = {Physics World Model},
  year    = {2026},
  url     = {https://github.com/integritynoble/Physics_World_Model},
}