GamePhysics-Counterfactuals-100K

A deterministic synthetic benchmark of 100,000 physics state transitions with matched hard counterfactual negatives.

Designed for:

• world-model evaluation
• physical reasoning
• game AI
• embodied AI
• action-conditioned prediction
• state-transition learning
• counterfactual consistency
• regression testing

Dataset summary

• 100,000 records
• 10 physics families
• 100,000 unique IDs
• 100,000 unique fingerprints
• deterministic analytical ground truth
• one hard counterfactual per example
• no copyrighted gameplay
• no real user data
• no PII
• no LLM-generated ground truth

Physics families

1. gravity_fall
2. projectile_motion
3. elastic_collision_1d
4. inelastic_collision_1d
5. friction_slide
6. bounce
7. constant_velocity_occlusion
8. support_removal
9. containment
10. momentum_transfer

Coverage:

{
  "gravity_fall": 10000,
  "projectile_motion": 10000,
  "elastic_collision_1d": 10000,
  "inelastic_collision_1d": 10000,
  "friction_slide": 10000,
  "bounce": 10000,
  "constant_velocity_occlusion": 10000,
  "support_removal": 10000,
  "containment": 10000,
  "momentum_transfer": 10000
}
Record format
Each record contains:
{
  "state_t0": { "...": "..." },
  "action": { "...": "..." },
  "expected_state_t1": { "...": "..." },
  "counterfactual_state_t1": { "...": "..." },
  "counterfactual_changed_field": "...",
  "label": {
    "expected_is_physically_consistent": true,
    "counterfactual_is_physically_consistent": false
  }
}

Counterfactual design
Every record includes:
- one mechanically valid next state
- one deliberately incorrect next state
- exactly one mutated output field
This makes the dataset useful for:
- binary physical-consistency evaluation
- preference learning
- hard-negative training
- reward-model evaluation
- world-model regression testing
Validation
The complete dataset passed an independent deterministic validator.
Results:
- 100,000 / 100,000 PASS
- 100,000 unique IDs
- 100,000 unique fingerprints
- independent next-state recomputation
- counterfactual single-field validation
- counterfactual invalidity verification
- provenance validation
Ground truth
Ground truth comes from analytical physics equations.
An LLM does not determine:
- positions
- velocities
- collision outcomes
- containment
- support behavior
- valid next states
- invalid next states
This keeps the benchmark deterministic and reproducible.
Important scope note
This is a state-transition physics benchmark, not a rendered video dataset.
It does not claim to contain:
- AAA gameplay
- real player traces
- photorealistic images
- real-world robotics trajectories
- full 3D rigid-body simulation
- fluid dynamics
Why counterfactuals?
Physical-AI systems often need more than positive examples.
They also need examples of states that look plausible but violate mechanics.
GamePhysics-Counterfactuals-100K provides both.
Commercial customization
Need this adapted to your world model or simulation stack?
Custom datasets can target:
- your state schema
- your action space
- your simulator
- longer horizons
- multi-object scenarios
- custom physics parameters
- curriculum difficulty
- known model failure modes
- private evaluation suites
Typical delivery:
Your state/action schema ? synthetic trajectories ? hard counterfactuals ? deterministic oracle ? validator ? QA report
Reproducibility
Default seed:
20260930
Generator and validator are included.
License
MIT.

---

## RegalFire — Custom / Private Dataset Work

RegalFire builds custom AI datasets, evaluation sets and data pipelines for:

- AI agents
- computer-use systems
- multimodal models
- world models
- RAG systems
- code agents
- enterprise AI

Available services include:

- synthetic data generation
- private evaluation datasets
- agent trajectories
- failure / recovery datasets
- multimodal RGB / segmentation / state-action data
- web data acquisition
- cleaning and deduplication
- structured dataset packaging
- continuous dataset production

For custom or private work:

**Email: ootiris@gmail.com**

Hugging Face: **RegalFire**