Handbook

MathGenesis governance

Mathematics and physics use different assurance semantics.

Updated 2026-07-26

Roles

Role	May read	May write	Must not do
Curator	approved sources, reference index	draft reference objects	self-validate or expose reference to learner
Learner	learner seed, curriculum, own staging	discovery staging	read reference or self-promote
Falsifier	candidate plus test tools	falsifier reports, rejected cases	repair before completing falsification
Validator	candidate, receipts, source bundle	bounded validation report	generate the candidate under review
Librarian	validated objects and ontology	ontology-patch proposal	silently merge or rename meaning
Promotion owner	all gate artifacts	promotion decision	waive blocker evidence through aggregate score
Historian	frozen discovery plus reference	historian-only equivalence report	leak modern aliases into learner snapshot
Release auditor	complete batch and repository index	read-only release report	write files or upgrade maturity

Truth boundaries

Mathematics and physics use different assurance semantics.

Mathematics

draft → source_grounded → bounded_tested
      → symbolically/numerically/formally_verified

Only a proof-kernel receipt supports formally_verified. Numerical and finite tests remain separate.

Physics

hypothesis → dimensionally_valid → fit_to_training_data
           → out_of_sample_validated → intervention_supported
           → regime_bounded → replicated

Solving equations correctly validates a computation, not the physical model.

Partition controls

Use separate storage roots and retrieval indexes for reference and learner-visible partitions.
Issue distinct credentials to learner and historian services.
Log every retrieval with snapshot and object IDs.
Keep hidden evaluator payloads outside learner prompts and tool responses.
Freeze source, model, prompt, and library hashes per generation.
Treat cross-session caches as possible contamination channels.

Promotion rule

Promotion is permitted only when all hard gates pass. Utility affects retention among valid candidates; it never turns a false or contaminated object into an accepted one.

Change rule

Meaning changes require a new semantic version. Breaking changes require a migration record. Never reuse an object ID for a materially different definition, theorem statement, model regime, or visibility policy.

Rejection policy

Retain useful failures:

falsified conjectures with minimal counterexamples;
invalid proofs with the first unsound step;
overfit physical models and held-out failures;
duplicate concepts and equivalence mappings;
contaminated clean-room episodes;
evaluator exploits and repaired evaluator specifications.

Rejected material trains falsifiers and curriculum designers, not the trusted knowledge generator.

Novelty claims

Use only:

new_to_agent;
new_to_snapshot;
new_to_local_reference;
potentially_novel_requires_external_review.

A claim of new mathematics requires external literature search, equivalence analysis, reproducible proof or computation, and expert review outside the autonomous loop.

Forge Intelligence