advantage_id	title	claim	proof	current_strength	durability	weakness	falsification_test	next_investment	evidence_routes
ADV-01	Evidence boundaries are product behavior	Measured, contextual, modeled, missing, failed and prohibited states remain distinct on public objects.	typed evidence states | claim ceilings | explicit abstentions | empty registries remain visible	PROVEN_PRODUCT_CAPABILITY	HIGH_IF_GOVERNANCE_REMAINS_ENFORCED	The advantage disappears if future releases collapse missingness into scores or hide failed runs.	Audit every new result route for source, task, execution state, prohibited uses and missing target-context outcome.	Make the reliability registry a required release gate.	/resource/claim_boundary | /resource/models/ | /resource/registry
ADV-02	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.	four disease routes | DMD process graph | cell-context registry | 21 governed DMD decision contracts	PROVEN_RESOURCE_CAPABILITY	MEDIUM_UNTIL_PROPRIETARY_PERTURBATION_TRUTH_EXISTS	Current DMD assets are rich context but do not contain matched candidate-conditioned DMD response truth.	Verify that every disease claim names the exact source context and inferential unit.	Acquire a matched DMD myogenic perturbation pilot with independent biological units.	/resource/diseases/ | /resource/process/ | /resource/cell-browser/
ADV-03	Research objects form a closed loop	Dataset, ModelRun, Study, Prediction and Outcome share identities and legal handoffs.	Dataset Cards | immutable ModelRun releases | Study Card planner | prospective lifecycle registry	STRUCTURALLY_IMPLEMENTED_NO_RETURNED_OUTCOMES	HIGH_AFTER_FIRST_PROSPECTIVE_OUTCOME_RETURN	The outcome registry is intentionally empty, so the loop is structurally complete but not yet empirically exercised.	A future outcome must trace back to its frozen study, prediction, data and decision rule without manual reconstruction.	Register and return the first prospective null, positive, toxic or failed outcome.	/resource/datasets/ | /resource/models/ | /resource/planner | /resource/registry
ADV-04	Simple baselines are permanent	Advanced models cannot advance by reputation, scale or one aggregate metric.	same-coverage baseline rule | frozen G0-G7 tasks | six-level evaluation ladder | GEARS and scGPT failures retained	PROVEN_METHOD_GOVERNANCE	HIGH_IF_BASELINES_REMAIN_MANDATORY	A baseline firewall protects decisions but does not create missing disease outcome.	Re-run each advanced model against identical data, split, genes, seeds and metric code.	Add perturbation-specific and population baselines as new output types become executable.	/resource/technology-methods/ | /resource/benchmarks/ | /resource/models/
ADV-05	Negative results remain reusable assets	Failed, no-go, stopped, null and abstained states are released with the same identity discipline as positive results.	six ModelRun release pages | reproduction receipts | null cross-disease results | prohibited-use fields	PROVEN_PRODUCT_CAPABILITY	HIGH_IF_SELECTION_BIAS_IS_RESISTED	Negative computational results do not replace prospective biological validation.	Compare public registries against the complete execution ledger and require every terminal run to be represented.	Attach resource-cost and failure-taxonomy fields to future ModelRuns.	/resource/models/ | /resource/model-run/mrun-gears-0-1-2-five-seed-20260713/ | /resource/diseases/compare/