protocolcompany/physics-verification-corpus

---
license: mit
task_categories:
- robotics
- reinforcement-learning
tags:
- physics
- simulation
- verification
- monte-carlo
- autonomous-systems
- robotics
- defense
- embodied-reasoning
size_categories:
- 100M<n<1B
---

# Physics Verification Corpus — 30 Companies

**130+ trajectory datasets. 30 companies. ~43 GB of deterministic physics.**

Generated from ~200 KB of hand-rolled Rust (no external physics libraries) on a single Apple M4 Pro. Every trajectory is SHA-256 sealed, anomaly-labeled, and 1000 Hz kinematic.

This corpus stress-tests the autonomous systems of 30 companies across robotics, defense, aerospace, autonomous vehicles, and eVTOL — exposing failure modes that simulation platforms (Isaac Sim, AirSim, CARLA) systematically miss.

---

## Companies & Domains

### Robotics (Round 1 — Live Fire Matrix)
| Company | Scenarios | Key Failure Modes |
|---------|-----------|-------------------|
| **Boston Dynamics** | Joint torsion, sim-to-real friction, thermal saturation, LBM inference | 360-degree joint shear, friction coefficient collapse |
| **Figure AI** | Payload shear, battery sag, tactile hallucination, kinematic resonance | Dynamic payload CG shift, haptic sensor ghosting |
| **Agility Robotics** | Digitigrade friction, elastomer latency, AMR docking, payload resonance | Toe-pad delamination, warehouse docking slip |
| **Skild AI** | Morphology thermal, gear backlash, sim mass distribution, compute starvation | Foundation model hallucination under thermal load |
| **Shield AI** | V-BAT crosswind, hivemind VIO drift, swarm wake, MEMS IMU resonance | Multi-agent visual-inertial odometry divergence |

### Autonomous Vehicles & eVTOL (Round 1)
| Company | Scenarios | Key Failure Modes |
|---------|-----------|-------------------|
| **Wayve** | Black ice hallucination, brake fade, hydroplaning, suspension asymmetry | End-to-end vision model failure on low-friction |
| **Joby Aviation** | Wake starvation, thermal throttle, acoustic resonance, actuator phase lag | Transition flight wake collapse, HOGE thermal |
| **Anduril** | VTOL transition, EKF divergence, sensor fusion, optics distortion | Lattice sensor fusion failure at Mach transition |

### Defense Primes (Live Fire Matrix Primes)
| Company | Systems Audited | Result |
|---------|----------------|--------|
| **Northrop Grumman** | Autonomous wingman, stealth thermal, UUV pressure, optical salt | 100% catastrophic failure |
| **Boeing Defense** | MQ-25 wake, CCA EMP, landing gear hysteresis, vision deck glare | 100% catastrophic failure |
| **General Dynamics** | Blast overpressure, radar mud attenuation, track pin galling, hydraulic stiction | 100% catastrophic failure |
| **BAE Systems** | AMPV suspension, EW drone swarm, gun barrel warp, armor spall | 100% catastrophic failure |
| **L3Harris** | Gimbal resonance, RF mesh packet storm, thermal sight, GPS spoofing | 100% catastrophic failure |
| **Huntington Ingalls** | USV hull slam, sonar thermocline, propeller cavitation, radar sea clutter | 100% catastrophic failure |
| **Textron Systems** | Rip-Saw brake fade, Aerosonde icing, CUES jamming, Cottonmouth harmonic | 100% catastrophic failure |
| **Kratos Defense** | Valkyrie flutter, Mako plasma blackout, Air Wolf collision, BTT laser | 100% catastrophic failure |
| **Leidos** | Sea Hunter thermal, cognitive radar spoof, cargo drone VRS, C2 BGP flap | 100% catastrophic failure |
| **AeroVironment** | Switchblade aliasing, Puma solar flare, Juggernaut CG shift, terminal dive shadow | 100% catastrophic failure |

### Robotics & AV (Round 2)
| Company | Scenarios | Key Failure Modes |
|---------|-----------|-------------------|
| **Tesla** | FSD hydroplaning, optical glare, Optimus tribology, harmonic slosh | Vision-only AV failure in adverse conditions |
| **1X Technologies** | Tendon snap, cable stretch backlash, ankle inversion, battery thermal | Cable-driven actuator mechanical failure |
| **Sanctuary AI** | Haptic chatter, hydraulic compressibility, thermal expansion, asymmetric drop | Dexterous manipulation physics gaps |
| **Unitree Robotics** | Metallurgic shear, actuator backlash gait, thermal sink, vibration chatter | Low-cost actuator failure cascades |
| **Archer Aviation** | Midnight PIO, rotor icing, battery sag, acoustic feedback | eVTOL transition and thermal failure |
| **Apptronik** | ZMP fall, gear galling, dynamic wind shear, thermal sensor starvation | Bipedal stability boundary collapse |
| **Skydio** | Optical shear, IMU acoustic resonance, thermal lens warp, VSLAM smoke | Autonomous drone navigation in degraded vis |
| **ANYbotics** | Thermal seal friction, IP67 heat soak, slip ring vibration, LiDAR refraction | Industrial inspection robot thermal limits |
| **Aurora Innovation** | Thermal outgassing, tire casing hysteresis, Doppler rain scatter, pneumatic resonance | AV sensor and mechanical degradation |
| **Waabi** | Black ice hallucination, trailer whip, air brake lag, sensor mud occlusion | Sim-to-real gap in trucking autonomy |

### Original Corpus (Pre-Matrix)
| Company | Datasets | Size |
|---------|----------|------|
| **Ghost Robotics** | Recoil, shale, subterranean drift, thermal (4.8M trajectories) | ~2.6 GB |
| **Amazon / Blue Origin** | FAR resmimic, Proteus hydraulic, lunar regolith, transonic (4.8M trajectories) | ~2.7 GB |
| **Elon Corpus** (SpaceX/Tesla/Boring/Neuralink/Starlink) | Starship catch, FSD, Boring thermal, Neuralink impedance, Kessler (6M trajectories) | ~3.4 GB |
| **DoD Hypersonic** | HGV Mach 20 plasma (1.2M trajectories) | ~751 MB |
| **Lockheed Martin** | F-35 transonic (1.2M trajectories) | ~690 MB |
| **US Navy** | Carrier PLM (1.2M trajectories) | ~631 MB |
| **Raytheon** | MALD EW drift (1.2M trajectories) | ~633 MB |

---

## Data Schema

Every trajectory follows this structure:

```json
{
"id": "unique_hex_id",
"type": "physics_type",
"scenario": "condition_description",
"steps": 1000,
"score": { "...domain-specific metrics..." },
"proof_hash": "sha256_hex",
"reasoning_context": {
"anomaly_type": "FAILURE_MODE_NAME",
"is_anomaly": true,
"snapshot": { "failure": "description" }
},
"data": [ "...timestep arrays..." ]
}
```

Every trajectory is **anomaly-labeled** with `reasoning_context`, making each record a training-ready sample for sim-to-real transfer learning.

---

## How to Verify

Each trajectory contains a `proof_hash` field (SHA-256). The proof chain is deterministic — given the same seed, the same physics, and the same integration step, you get the same hash. No stochastic noise. No random seeds. Pure math.

```python
import json, hashlib

with open("ghost_robotics/ghost_robotics_recoil_1_2M_sealed.json") as f:
data = json.load(f)

# Check any trajectory
traj = data["trajectories"][0]
print(f"ID: {traj['id']}")
print(f"Proof: {traj['proof_hash']}")
print(f"Anomaly: {traj['reasoning_context']['anomaly_type']}")
print(f"Survived: {traj['score'].get('survived', 'N/A')}")
```

---

## Generation

- **Engine:** G^G (genesis_core) — hand-rolled Rust, Euler/symplectic integration at 1000 Hz
- **Machine:** Apple M4 Pro (14-core CPU, 20-core GPU, 24 GB unified memory)
- **Physics:** No external libraries. Every force law is written from first principles.
- **Sealing:** SHA-256 per-trajectory, proof-chained across datasets

---

## Source

- **GitHub:** [github.com/aijesusbro/Spectrum](https://github.com/aijesusbro/Spectrum)
- **Website:** [aijesusbro.com](https://aijesusbro.com)
- **Organization:** Protocol Company

---

## License

MIT