Forge Intelligence architecture alignment for PhysicsGenesis

Where the physics library belongs

The handbook is a generated static documentation surface. The writable source of truth is the forge-intelligence product repository. PhysicsGenesis should generate repository-native, reviewable artifacts and then allow the existing website build to publish selected documentation and wiki projections. It should not target generated HTML.

Existing architecture to reuse

Forge Intelligence already supplies:

• general and specific problem-class packs;
• per-pack policies with reasoning ladders, token budgets, and closed actions;
• one capability ↔ one LMeta unit ↔ one FI-local LCDL contract ↔ declared artifacts;
• typed artifact-envelope handoffs;
• multi-intelligence DAG composition with dependencies and fan-in;
• separate technical, autonomy, and evidence maturity axes;
• implemented allowlist gates;
• machine-canonical session records and generated human reports with a freeze gate;
• paired LMeta semantic and LCDL execution wiki projections;
• reusable reasoning, concept invention, experimentation, system modeling/simulation, creativity, and validation intelligences.

Forge LCDL contributes generic governed tasks such as schema extraction, routing, boolean and multi-label gates, contradiction scanning, problem decomposition, decision-pack planning, safe arithmetic planning, and reusable control-flow operators. PhysicsGenesis composes those primitives; it does not fork the runtime.

Gap this pack fills

The current generic knowledge-module representation is adequate for bounded changing facts but is too shallow for physics as a scientific knowledge system. Physics needs native distinctions among:

• quantities, dimensions, units, constants, and conventions;
• principles, empirical laws, constitutive laws, equations, derivations, approximations, and regimes;
• models, alternatives, parameters, observables, controls, and causal status;
• apparatus, protocols, measurements, calibration, uncertainty, datasets, and evidence roles;
• simulation models, numerical methods, residuals, convergence, invariants, and reproducibility;
• anomalies, falsification cases, rejected models, and replication;
• learner-visible material, hidden world truth, validator-only evaluators, and historian mappings;
• laboratory hazard and dual-use detail policy.

Runtime gap

Envelope execution, schema validity, and maturity status do not establish physics correctness. PhysicsGenesis therefore requires independent semantic validators and authenticated receipts for dimensions, derivations, numerical verification, measurement traceability, held-out prediction, interventions, replication, contamination, and safety.

Proposed problem classes

Non-negotiable separation

• The modern reference index must be inaccessible to the learner.
• Simulator truth and evaluator internals must be inaccessible to the learner.
• A validator must not be the candidate generator.
• A promotion owner must not waive failed hard gates through utility scoring.
• A historian operates only after generation freeze.
• Safety and epistemic judgments remain separate artifacts.
• Rejected artifacts are retained for falsifier and curriculum learning, not trusted positive knowledge.

Recommended rollout

1. Add native schemas and the compatibility wrapper.
2. Add partition-scoped storage and retrieval before any clean-room experiment.
3. Implement quantity normalization, model formalization, experiment design, simulation assessment, safety, falsification, and validation capabilities.
4. Ingest the first twelve reference modules.
5. Build and freeze Physica-0 worlds, seeds, and hidden evaluation maps.
6. Run learner generations with complete session and retrieval traces.
7. Add training-record generation only from policy-eligible frozen artifacts.
8. Publish documentation after LMeta/LCDL identity parity and release checks pass.