Forge Intelligence

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

  1. Select a mixed frontier curriculum.
  2. Run observation, measurement, intervention, modeling, and prediction episodes.
  3. Freeze the episode archive.
  4. Mine candidate inventions.
  5. Run independent falsification.
  6. Run independent validation.
  7. Compare with learner-visible ontology only.
  8. Run safety and contamination audits.
  9. Measure held-out prediction, compression, transfer, experiment efficiency, complexity, and brittleness.
  10. Apply human-gated promotion.
  11. Freeze the next learner snapshot.
  12. 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.