Forge Intelligence

intelligence_physics_model_assess_v1 v1

Purpose: Assess a unit-aware physical model separately from the correctness of computations performed with it.

Updated

Inputs

Required: model_candidate, unit_registry, training_observations, heldout_observations, assessment_policy.

Outputs

Required: model_assessment, status.

Substrate reuse

  • Generic tasks: extract_schema_from_text, contradiction_scan, llm_multi_label
  • Operators: seq, branch, fallback_chain, for_each, repeat, try_catch, until_ok

Verification policy

  • units and dimensions
  • frame and sign conventions
  • residual and convergence evidence
  • train/heldout separation
  • regime and omitted effects explicit

Failure modes

  • dimension mismatch
  • data leakage
  • overfit model
  • out-of-regime extrapolation
  • missing uncertainty decomposition

Runtime status

Proposal only. Repository-native implementation, fixtures, validators, maturity registration, and allowlist entry are required.