# NMD-VCell v2.2 fold-stability model-card addendum

Model: `safe_external_v1`  
Parent card: `55_v21_primary_model_card_addendum_20260717.md`  
Status: verified same-HepG2 sensitivity analysis

## What changed

The frozen primary model, features, target universe, endpoints and seed-20260717
estimate did not change. Nineteen prespecified balanced fold assignments were
added to measure fold-assignment and fold-local training variability.

Each of 20 realizations used 2,160 targets, five non-overlapping 432-target outer
folds, training-fold imputation/scaling, five-fold inner CV and the locked ridge
alpha grid. All 100 outer-fold fits selected alpha 1000. Three negative controls
were repeated in every realization.

## Stability results

| Metric | Frozen primary | 20-realization mean | Between-realization SD | Joint 95% sensitivity interval | Interpretation |
|---|---:|---:|---:|---:|---|
| RMSE difference vs zero | −0.006314 | −0.006292 | 0.000033 | −0.007352 to −0.005240 | stable |
| RMSE difference vs train mean | −0.000480 | −0.000466 | 0.000029 | −0.000702 to −0.000225 | stable but small |
| Raw-cosine difference vs train mean | +0.001637 | +0.001224 | 0.000275 | −0.000216 to +0.002498 | mixed |
| Perturbation-specific residual cosine | +0.04544 | +0.04528 | 0.00144 | +0.02965 to +0.06118 | stable secondary signal |

All 20 raw-cosine point estimates were positive, ranging from 0.000571 to
0.001707. Nevertheless, the hierarchical sensitivity interval crossed zero
because it incorporates target sampling as well as observed fold/training
variation. The frozen primary conditional P=0.00356 must therefore not be
described as fold-robust directional support.

## Updated permitted claim

Response-independent features produce a stable but modest reduction in average
same-HepG2 OOF response-vector error relative to both zero and the outer-training
mean across prespecified balanced fold assignments. Raw-direction point
estimates are consistently positive but not robust under the joint target and
fold/training sensitivity interval.

## Prohibited interpretations

- Do not call the model a strong target-specific directional predictor.
- Do not use the 20 seeds as independent biological replicates.
- Do not pool target rows across seeds as 43,200 independent observations.
- Do not replace the frozen primary estimate with the most favorable seed.
- Do not infer muscle/DMD response, mechanism, efficacy, safety or clinical use.

## Verification

The repeated analysis passed 120/120 independent checks. Fresh seed-20260717
folds reproduced exactly; summary-level numerical differences were below
`5e-08`. The accepted float32 target-metric reproduction tolerance and its
outcome-blind correction are documented in the v2.2 debug log.