Answers are grounded in this paper and cite their sources.
Validation study demonstrates persistent apparatus-descriptor irreducibility across synthetic grids, indicating that physics-informed descriptors fail pre-registered recovery criteria.
GMT v2.6 found descriptor-irreducible apparatus dependence inside its discovery grid: even a rich apparatus-descriptor basis did not recover the measured geometry-range score G₁ under out-of-apparatus validation. GMT v2.7 asks whether that irreducibility persists on an independent synthetic grid, behind an information firewall, when the descriptors are selected a priori from physics rather than fitted to the External targets. The study freezes before External G₁ generation: (i) a six-descriptor physics-informed basis constructed without target access; (ii) a Legacy reference, a flat ridge model used as the primary A/B/C decision model, and a hierarchical model retained only as a secondary robustness analysis; (iii) all standardisation statistics, penalties, thresholds, and decision rules; and (iv) an External grid of 12 apparatus classes across three range regimes, comprising 36 cells balanced across four configuration families. The External grid includes a compensated-az-force family absent from the Discovery panel. After the freeze, External G₁ is generated once using the unchanged v2.4/v2.5 Monte-Carlo machinery, and the frozen pipeline is evaluated without refitting, descriptor modification, class exclusion, clipping, or threshold adjustment. Results. On the External grid, the primary flat model attains global RMSE = 0.952 dex, global R² = +0.258, family-balanced RMSE = 0.903 dex, and worst-family RMSE = 1.205 dex. The model beats both the grand-mean total-variance reference (RMSE = 1.105 dex; R² = 0) and the Discovery-trained constant baseline (RMSE = 1.140 dex; R² = −0.065), showing that the frozen descriptors carry real transferable signal. However, performance remains far below the pre-registered recovery thresholds. No configuration family meets the joint family-level criterion, and both mechanical Test A and Test B fail. The frozen decision rule therefore returns Outcome C — persistent descriptor irreducibility. Outcome C is strictly scoped to the transferability of this frozen physics-informed basis, these frozen models and thresholds, and this one independent synthetic External grid. It is not a claim of universal or physically essential unpredictability, not proof that no alternative descriptor or model could transfer, and not a PASS certification: the sampled one-shot evaluation yields FAIL, not PASS, and no exact or interval guarantee is claimed. All apparatus are synthetic representative classes; none is HUST hardware, and no experimentally demonstrated gravitational manipulation is claimed. The accompanying supplement contains the frozen Master Specification, Stage 1–3B Python code, Discovery and External JSON data, RAW External G₁, evaluation results, stage reports, Figures 1–8, and SHA-256 integrity manifests.
No takes yet. Share an insight, caveat, or question.
Koji Okino (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: