Handbook
Physica-0 clean-room laboratory protocol
Can a learner invent reusable physical quantities, relations, models, measurement operations, experiment strategies, and invariants from minimal primitives and controlled problem pressure without access to modern physics…
Updated
Experimental question
Can a learner invent reusable physical quantities, relations, models, measurement operations, experiment strategies, and invariants from minimal primitives and controlled problem pressure without access to modern physics labels or solutions?
Isolation
- Use an audited learner model or explicitly label pretraining contamination risk.
- Disable reference retrieval and cross-session caches.
- Use separate learner, validator, safety, and historian credentials.
- Freeze prompt, model, seed, world, evaluator, and library hashes.
- Randomize superficial labels and constants while preserving structure.
Generation cycle
- Select a mixed frontier curriculum.
- Run observation, measurement, intervention, modeling, and prediction episodes.
- Freeze the episode archive.
- Mine candidate inventions.
- Run independent falsification.
- Run independent validation.
- Compare with learner-visible ontology only.
- Run safety and contamination audits.
- Measure held-out prediction, compression, transfer, experiment efficiency, complexity, and brittleness.
- Apply human-gated promotion.
- Freeze the next learner snapshot.
- Run historian comparison separately.
Required controls
- Baseline learner without library mining.
- Learner with episode memory but no concept promotion.
- Learner with concept promotion but no cross-domain transfer.
- Full PhysicsGenesis learner.
- Multiple random seeds and world parameterizations.
Success
A discovery counts only when it is operationally defined, improves held-out performance or compression, survives bounded falsification, has explicit regime and uncertainty, passes contamination and safety checks, and can be reproduced from the frozen episode record.
Non-successes worth retaining
- overfit laws;
- unmeasurable quantities;
- duplicate concepts;
- models with hidden frame dependence;
- invalid conservation claims;
- apparatus-confounded regularities;
- unsafe experiment proposals;
- evaluator exploits;
- contaminated rediscoveries.