Skip to content

Latest commit

 

History

History
38 lines (29 loc) · 1.98 KB

File metadata and controls

38 lines (29 loc) · 1.98 KB

Validation programme

TransferMod v1.2.1 treats empirical scope as an open question rather than an assumption.

Track Current evidence Interpretation
Learned operators the trained random-feature diffusion operator shows no point with relative global L2 error at or below 0.01 and decision error at or above 0.02 negative result for decoupling in this model/family; aggregate and decision degradation remain coupled
Higher-dimensional PDEs compact 2D/3D uniform and coarse-to-fine searches execute successfully implementation-feasibility result only; production solver scaling remains open
Bayesian Silent Risk conjugate drift–diffusion posterior cleanly separates posterior weighting, anchored fidelity, worst-case fidelity, and decision diameter successful implementation validation, not evidence of broad posterior robustness

Immediate confirmatory priority

Run the frozen protocol in LEARNED_OPERATOR_PROTOCOL.md on externally trained operators that were neither designed nor trained for TransferMod.

The first confirmatory study should include:

  • publicly released FNO and DeepONet checkpoints;
  • model and checkpoint hashes fixed before analysis;
  • predeclared aggregate metrics, quantities of interest, thresholds, families, and search budgets;
  • complete reporting of coupled, decoupled, and inconclusive outcomes;
  • no post-result substitution of perturbation families in the confirmatory analysis.

Remaining scale-up work

  • external learned-operator replication;
  • solver-coupled two- and three-dimensional PDE studies with measured hardware cost;
  • non-conjugate, multimodal, and MCMC posterior Silent Risk;
  • preregistered comparative studies across models, metrics, quantities, and information bases.

The scope question is now explicit: under what conditions do aggregate and decision fidelity decouple in real trained systems? A negative answer for a particular model–metric–quantity–family combination is a first-class result.