Handbook
PhysicsGenesis governance
The branches are not interchangeable. A formal derivation can coexist with no empirical validation. A replicated empirical model may still be effective rather than fundamental.
Updated
Roles
| Role | May read | May write | Must not do |
|---|---|---|---|
| Curator | approved sources and reference index | draft reference objects | invent evidence or expose reference to learner |
| Metrologist | quantity, unit, calibration, and uncertainty sources | normalization and uncertainty records | silently alter quantity meaning or significant digits |
| Theorist | approved models and mathematical foundation | draft theory/model cards | claim empirical validity from derivation alone |
| Experimentalist | candidate models, apparatus, constraints | experiment and protocol cards | rewrite post-hoc analysis as preplanned |
| Simulator | model card, code, configuration | simulation receipts | validate the physical model by solver success |
| Learner | learner seed, curriculum, own staging | discovery staging | read reference, evaluator truth, or self-promote |
| Falsifier | candidate and test interfaces | falsifier reports and counterexamples | repair before independent attack completes |
| Validator | candidate, receipts, sources, data | bounded validation report | generate the candidate under review |
| Safety reviewer | task and intended-use context | safety report | treat safety as truth evidence |
| Librarian | validated objects and permitted ontology | ontology patch | silently merge different regimes or operational meanings |
| Promotion owner | all gate artifacts | promotion decision | override blockers with aggregate utility |
| Historian | frozen discovery plus modern reference | historian-only comparison | leak aliases or targets back to learner |
| Release auditor | complete batch and repository snapshot | read-only release report | mutate files or upgrade maturity |
Physics evidence ladder
draft → source_grounded → dimensionally_valid → mathematically_consistent
→ formally_derived / computationally_verified
→ fit_to_training_data → out_of_sample_validated
→ intervention_supported → regime_bounded → replicated
The branches are not interchangeable. A formal derivation can coexist with no empirical validation. A replicated empirical model may still be effective rather than fundamental.
Partition controls
- Separate storage roots, retrieval indexes, credentials, and caches for reference and learner-visible partitions.
- Trace every retrieval with snapshot and object IDs.
- Freeze model, prompt, source, library, world, evaluator, and code hashes per generation.
- Store world ground truth and target maps outside learner-visible artifacts.
- Never direct-write from a generator to
promoted_discovery.
Promotion rule
All hard gates must pass. Utility chooses among valid candidates; it does not make a false, contaminated, unsafe, or unsupported claim acceptable.
Versioning rule
A changed quantity definition, unit conversion, frame convention, model assumption, equation, apparatus protocol, dataset split, or validity regime requires a semantic version change. Breaking changes require migration and supersession records.
Rejection policy
Retain:
- refuted laws and minimal counterexamples;
- overfit models and held-out failures;
- invalid derivations and first unsupported step;
- numerical instability and non-convergence traces;
- calibration failures and null results;
- unsafe procedural variants;
- duplicate concepts and regime mappings;
- contaminated episodes and evaluator exploits.
Use them for falsifier, repair, and curriculum training—not as trusted positive knowledge.
Novelty labels
Only use new_to_agent, new_to_snapshot, new_to_local_reference, or potentially_novel_requires_external_review. A claim of new physics requires external search, reproducible evidence, equivalence analysis, and expert review beyond the autonomous loop.