Reliability & Advantage Center · assessed 2026-08-16
Reliable where evidence is measured. Abstains where truth is missing.
NMD-VCell does not publish one trust score. It separates engineering integrity, evidence synthesis, task-level prediction and clinical validity because success in one layer cannot upgrade another.
Five defensible advantages
An advantage needs proof, a weakness and a falsification test.
These claims describe implemented product behavior. They do not claim a superior DMD predictor.
Evidence boundaries are product behavior
Measured, contextual, modeled, missing, failed and prohibited states remain distinct on public objects.
- Proof
- typed evidence states · claim ceilings · explicit abstentions · empty registries remain visible
- Weakness
- The advantage disappears if future releases collapse missingness into scores or hide failed runs.
- Falsification test
- Audit every new result route for source, task, execution state, prohibited uses and missing target-context outcome.
- Next investment
- Make the reliability registry a required release gate.
Neuromuscular context is first-class
Disease, cell state, perturbation, time and phenotype are represented as the scientific unit rather than collapsed into a gene score.
- Proof
- four disease routes · DMD process graph · cell-context registry · 21 governed DMD decision contracts
- Weakness
- Current DMD assets are rich context but do not contain matched candidate-conditioned DMD response truth.
- Falsification test
- Verify that every disease claim names the exact source context and inferential unit.
- Next investment
- Acquire a matched DMD myogenic perturbation pilot with independent biological units.
Research objects form a closed loop
Dataset, ModelRun, Study, Prediction and Outcome share identities and legal handoffs.
- Proof
- Dataset Cards · immutable ModelRun releases · Study Card planner · prospective lifecycle registry
- Weakness
- The outcome registry is intentionally empty, so the loop is structurally complete but not yet empirically exercised.
- Falsification test
- A future outcome must trace back to its frozen study, prediction, data and decision rule without manual reconstruction.
- Next investment
- Register and return the first prospective null, positive, toxic or failed outcome.
Simple baselines are permanent
Advanced models cannot advance by reputation, scale or one aggregate metric.
- Proof
- same-coverage baseline rule · frozen G0-G7 tasks · six-level evaluation ladder · GEARS and scGPT failures retained
- Weakness
- A baseline firewall protects decisions but does not create missing disease outcome.
- Falsification test
- Re-run each advanced model against identical data, split, genes, seeds and metric code.
- Next investment
- Add perturbation-specific and population baselines as new output types become executable.
Negative results remain reusable assets
Failed, no-go, stopped, null and abstained states are released with the same identity discipline as positive results.
- Proof
- six ModelRun release pages · reproduction receipts · null cross-disease results · prohibited-use fields
- Weakness
- Negative computational results do not replace prospective biological validation.
- Falsification test
- Compare public registries against the complete execution ledger and require every terminal run to be represented.
- Next investment
- Attach resource-cost and failure-taxonomy fields to future ModelRuns.
Reliability matrix
Six domains; six different claim ceilings.
No composite reliability score is emitted because engineering integrity, evidence synthesis, perturbation prediction and clinical validity are non-interchangeable claims.
Software & release integrity
Build, route, schema and cross-surface gates pass locally.
Current evidencefull vinext production build PASS · 51 rendered contract tests · 246 multidisease checks · 24 cross-surface checks · checksum-addressed release objects
- Permitted use
- Rebuild, audit and navigate the released research resource.
- Limitation
- Anonymous command-line production QA is blocked by the existing edge HTTP 403 policy.
- Do not infer
- Engineering integrity does not validate a biological prediction.
Identity, provenance & data contracts
Identifiers, sources, contexts, nulls and transformations are explicit.
Current evidence17,921 released gene records · HGNC alias map · Dataset Cards · typed API · checksums and release identities
- Permitted use
- Identity resolution, evidence lookup and reproducible data intake.
- Limitation
- A valid data object can still be unsuitable for a particular prediction task.
- Do not infer
- Presence in the resource is not disease relevance or efficacy.
Disease evidence synthesis
Source-linked context is strong enough for hypothesis and study design.
Current evidenceDMD measured-evidence release · four disease modules · cell-context map · process and evidence graphs
- Permitted use
- Mechanism navigation, gap analysis and prospective study design.
- Limitation
- Observational and correction context does not establish candidate-conditioned causal response.
- Do not infer
- Do not treat integrated context as target efficacy or causal proof.
Same-context perturbation benchmark
Ridge slightly improves mean RMSE, but response direction is unsupported.
Current evidenceMRUN-RIDGE-SAFE-2.3-G0-REPEATED-FOLD · RMSE 0.1175 vs train mean 0.1180 · strict outcome gate 0/16
- Permitted use
- Permanent same-assay comparator and leakage diagnostic.
- Limitation
- The output is a 2,000-feature aggregate HepG2 mean vector, not a cell population or DMD response.
- Do not infer
- No directional transfer, DMD prediction or functional effect.
Cross-context & DMD response prediction
No model has passed a matched DMD candidate-response task.
Current evidenceGEARS failed five frozen splits · scGPT lost to same-coverage baselines · TxPert calibration pending · direct DMD candidate outcome 0/21
- Permitted use
- Method research, task design and governed abstention.
- Limitation
- The required donor- and cell-state-matched perturbation outcome matrix does not exist in the release.
- Do not infer
- No candidate ranking, efficacy prediction or virtual DMD response population.
Therapeutic, patient & clinical use
No clinical calibration, prospective outcome set or authorization exists.
Current evidencezero calibrated DMD models · zero prospective predictions · zero measured outcomes · DOI and institutional licences pending
- Permitted use
- None beyond research hypothesis and study-design support.
- Limitation
- The platform is a research resource, not a medical device or treatment recommender.
- Do not infer
- No diagnosis, prognosis, treatment selection, dosing or patient-level digital twin.
Machine-checked receipt
Engineering confidence has a reproducible trail.
These checks protect route, schema and release integrity. They do not validate biological predictions.
- Full build, scientific validators, rendered contracts and lint
- EXTERNAL POST BUILD RECEIPT REQUIRED
- A source bundle cannot prove its own later deployment.
- Content commit ·
91e7e8ff973d - https://nmdvcell.com/resource/
- Automated edge QA · BLOCKED HTTP 403
Moat-building plan
The current moat is trust infrastructure. The future moat is proprietary DMD perturbation outcome.
Each stage names the object that must exist before the advantage can strengthen.
Trust infrastructure
Typed evidence, model and reliability contracts with negative-result preservation.
- Acceptance gate
- Every claim has provenance, execution state, failure state, ceiling and next gate.
- No shortcut
- Do not replace explicit states with a synthetic trust score.
Matched DMD perturbation pilot
Candidate-conditioned DMD myogenic perturbation outcomes with independent biological units, molecular and functional readouts.
- Acceptance gate
- Frozen estimand, controls, donor/culture units, QC, toxicity and complete outcome return.
- No shortcut
- Do not substitute healthy myoblast, HepG2 or unperturbed DMD context.
Donor-resolved benchmark
Leakage-safe donor, state and perturbation holdouts with permanent simple baselines.
- Acceptance gate
- Perturbation-specific, population, calibration and functional layers pass prospectively.
- No shortcut
- Cells cannot be treated as independent donors and mean RMSE cannot be the only metric.
Prospective learning loop
Registered predictions whose positive, null, toxic and failed outcomes return to the same evidence graph.
- Acceptance gate
- Immutable preregistration, blinded evaluation, calibration and independent replication.
- No shortcut
- Retrospective fit cannot be relabeled as prospective validation.
Update rule
A new page cannot upgrade a reliability state.
A reliability state changes only when its named next gate produces a versioned object; a new page, model name or external citation cannot upgrade it.
This registry audits current product and evidence reliability. It is not a certification, medical-device assessment, therapeutic recommendation or substitute for prospective biological and clinical validation.