| 1 | Verified cohort, frozen split and no-RNA baseline | Replace the recalled count with the verified data; distinguish endpoint fitting from a connected kinetic model. | -0.001 | All 274 matched lines retained. Test identities unchanged; preprocessing and supervised filtering are local to training folds. |
| 2 | Connect RNA to damage/repair/injury kinetics | Train through exact state integration instead of an isolated mSF2 regression head. Sweep L2 to control overfitting. | 0.276 | Large training–validation gap remains; increased capacity is not evidence of better generalization. |
| 3 | Test hidden-layer capacity | The first models can overfit hundreds of inputs; test one/two bottleneck layers with shrinkage. | 0.276 | Large training–validation gap remains; increased capacity is not evidence of better generalization. |
| 4 | Train for identity-line agreement | MSE favors compressed predictions. Add concordance loss inside training rather than stretching outputs after evaluation. | 0.282 | Large training–validation gap remains; increased capacity is not evidence of better generalization. |
| 5 | Fold-local RNA feature selection | Reduce irrelevant inputs without using held-out labels; gene ranks are re-estimated inside every training fold. | 0.282 | Large training–validation gap remains; increased capacity is not evidence of better generalization. |
| 6 | Pathway bottlenecks and learned inter-module edges | Compare all-gene connections with the inherited gene/function priors and learn recurrent message-passing edges. | 0.282 | Large training–validation gap remains; increased capacity is not evidence of better generalization. |
| 7 | Nonlinear injury and time-resolution tests | Allow pairwise misrepair in the same kinetic injury integral; compare with linear damage and a no-stress ablation. | 0.282 | Large training–validation gap remains; increased capacity is not evidence of better generalization. |
| 8 | Robustness and smooth nonlinear gene mapping | Test robustness to RNA noise and outlier loss, and a radial-basis encoder; do not delete difficult cell lines. | 0.316 | Large training–validation gap remains; increased capacity is not evidence of better generalization. |
| 9 | Five-seed ensemble and selection lock | Reduce initialization variance without selecting the seed that happens to fit a test set. | 0.295 | Five seeds averaged equally; ensemble selection-stage CCC 0.295. No best-seed picking. |
| 10 | Nested adaptive search, locked final fit and reused-test audit | Separate model selection from evaluation; verify whether a strong training fit transfers to unseen cell identities. | Nested 0.155 / test 0.197 | Nested CCC 0.155, test CCC 0.197. Test MAE 0.1411 is worse than V1.4. Kept the locked model and all results. |