# Frozen scientific identity and claim ladder

## Identity

- Submission category: regular **NAR Data Resources and Analyses**.
- Resource identity: **NMD-VCell Audit**, an auditable benchmark and evidence resource for direction-aware evaluation of perturbation-response models, with DMD-guided hypothesis triage as a bounded use case.
- Benchmark domain: CRISPR perturbations measured in HepG2.
- Disease layer: an integrated DMD direction prior assembled from heterogeneous observational sources, plus source-specific effects and sensitivity views.
- Prohibited identity: a validated DMD virtual cell, a muscle perturbation predictor, a therapeutic-target discovery engine, or a pan-NMD predictor.

## Claim ladder

| Level | Claim | Evidence requirement | Current status |
|---|---|---|---|
| L1 | Aggregate perturbation deltas contain technically repeatable structure | Target-level split-half/batch or biological-replicate agreement with cells used only to estimate target deltas | Partly supported: cell/batch median cosine ≈0.33; guide-pair median cosine ≈0.006 in the limited auditable subset |
| L2 | A model generalizes internally to held-out perturbation identities in the same HepG2 dataset | Frozen outer split, train-only tuning, raw-direction primary endpoint, no-effect baselines, coverage accounting and leakage sensitivity | Narrowly supported for RMSE and residual structure; raw direction does not exceed train mean |
| L3 | A model generalizes across cell types or datasets | Untouched external perturbation-response dataset with matched endpoint and prespecified evaluation | Not supported |
| L4 | A model captures DMD-relevant perturbation direction in muscle | Disease-relevant muscle/myotube/organoid perturbation transcriptomics with independent replication | Not supported |
| L5 | A candidate has therapeutic efficacy or clinical utility | Prospective mechanistic, preclinical and clinical evidence | Not supported |

## Endpoint hierarchy frozen for future work

1. Primary: raw response-direction cosine and sign concordance at the target level.
2. Co-primary reliability context: replicate/guide/batch agreement and target-level uncertainty.
3. Secondary: RMSE against zero-delta and train-mean baselines.
4. Diagnostic only: common-effect-adjusted residual cosine.
5. Triage only: DMD-prior rescue-like scores, evidence class and action priority.

Passing a lower item cannot compensate for failure of a higher item. Residual cosine alone cannot establish biological response direction. Candidate priority cannot establish model validity or treatment benefit.

## Route boundary

The selected journal category is not being changed. The scientific narrative is being narrowed so that the regular NAR Data Resources and Analyses submission is evaluated as a reusable, versioned audit resource. A disease-validation route would require new muscle/DMD perturbation data and is outside the current evidence package.