Deploying industrial world models across different machines requires transfer across changes in dynamics, sensing interfaces, sampling regimes, and control units. This work investigates a Schema‑Adaptive Action‑Conditioned Joint‑Embedding Predictive Architecture (SAAC‑JEPA) for CNC dynamics, where the source machine provides 17 canonical sensor channels while the target shares only 10 of them. The evaluation protocol includes:
- Group‑disjoint source splits;
- Source‑only normalization;
- Held‑out self‑supervised validation for early stopping;
- Unit audits;
- A sealed target test after model locking.
Across five random seeds, JEPA pre‑training yields no clean‑source forecasting gain: training from scratch achieves RMSE $0.811\pm0.022$, and a pretrained‑body model reaches $0.813\pm0.022$. A source‑only search over 20 candidates, followed by seven‑seed stability checks, selects a schema‑consistent action‑conditioned JEPA. In the confirmatory target pass, the locked model attains zero‑shot RMSE $0.546$, $R^2=0.012$, and negative log‑likelihood (NLL) $0.52$, outperforming a persistence baseline but lagging behind RevIN‑equipped PatchTST and iTransformer baselines ($0.503$ and $0.498$ respectively).
A paired ablation shows that adding RevIN to the same architecture improves RMSE to $0.495\pm0.004$ (three seeds) but severely degrades target calibration (NLL $20.6$) on stationary context windows. A pre‑lock adaptation sweep further reduces RMSE to $0.520$ with limited target support.
These findings demonstrate that source‑domain forecasting accuracy alone is insufficient to assess industrial predictive representations, and that cross‑machine adaptation under partial sensor overlap constitutes a distinct evaluation axis.
Review