Retain as a mandatory comparator in every future task.

Keep fold-specific fitting and reporting.

Preserve as the interpretable model floor.

Publish the exact aggregation recipe when source-level raw objects are available.

Add training-only perturbation similarity and a leakage-safe neighbour baseline.

Evaluate an openly reproducible model only after exact input and output parity is established.

Keep this as the lowest generalization tier, not as disease transfer.

Acquire donor-resolved muscle/DMD perturbation data and lock donor-disjoint splits.

Create accession-disjoint training and test objects after modality harmonization.

Define stage labels and hold out complete disease stages, never random cells.

Generate or import matched muscle-context perturbation outcomes.

Sample the same candidate perturbations across matched differentiation timepoints.

Rebind MAE only after the official 2026 expression scale and prediction unit are known.

Retain the current negative result: raw-direction support is mixed.

Add up/down DEG precision-recall and threshold sensitivity using cell-eval-compatible outputs.

Re-execute with official 2026 AnnData inputs and preserve raw plus scale-audit outputs.

Gate Wasserstein/MMD or cell-eval distribution metrics on genuine cell-level predictions.

Score prespecified program changes against held-out observed responses.

Evaluate only when model outputs represent cell-level distributions.

Extend uncertainty to donor, dataset and disease-stage components.

Use as context qualification, never as perturbation truth.

Expand candidate coverage and measure multiple muscle functions.

Use as a disease-state benchmark, not as candidate causal evidence.

Register predictions before assays and bind them to immutable outcome definitions.

Prespecify one functional primary endpoint and bounded secondary endpoints per study.

Keep clinical utility locked until independent prospective validation.