Forge Intelligence

Physics simulation assess v1

Purpose: Assess numerical reproducibility, residuals, convergence, invariants, comparisons, and uncertainty without validating the model by implication.

Updated

Input

Required fields: model_card, configuration, run_results. Additional bounded context may be supplied according to contract.json.

Output

{ "simulation_validation": <object> } where the payload validates against schemas/physicsgenesis/simulation-validation.schema.json.

Substrate reuse

Compose Forge LCDL tasks: contradiction_scan, formula_eval_safe, llm_boolean_gate. Use standard operators for guarded sequencing, fallback, and bounded repair.

Failure modes

  • missing code version
  • no refinement evidence
  • residual failure
  • model error conflated with numerical error

Verification

  • JSON Schema
  • replay metadata
  • convergence sequence
  • invariant and residual results

The task never self-promotes, invents receipts, or upgrades empirical status without authenticated evidence.