{
  "registry_schema": "nmd-vcell-benchmark-task-registry/2.0",
  "generated_at": "2026-09-09",
  "comparison_rule": "Models are compared only within the same task, split, feature space, endpoint and independent-unit contract.",
  "tasks": [
    {
      "task_id": "G0",
      "label": "Same-context target holdout",
      "state": "EXECUTED_LIMITED",
      "split_contract": "repeated-fold-v2.2 target holdout",
      "feature_space": "frozen 2,000-feature response vector",
      "endpoint": "Held-out 2,000-feature mean response vector",
      "independent_unit": "held-out perturbation target; cells do not increase the target denominator",
      "permanent_baselines": [
        "zero change",
        "train mean",
        "ridge"
      ],
      "primary_metrics": [
        "RMSE",
        "raw cosine",
        "residual cosine",
        "strict 16-outcome gate"
      ],
      "eligibility_gate": "Released same-assay target aggregates",
      "current_result": "Small average-error gain; strict gate 0/16 and direction unreliable",
      "decision_blocker": "No external cell-level or disease-context truth",
      "route": "/resource/benchmarks/core/"
    },
    {
      "task_id": "G1",
      "label": "External unseen perturbation",
      "state": "EXECUTED_PARTIAL_UNSUPPORTED",
      "split_contract": "new perturbation in an external aggregate object",
      "feature_space": "source-defined; exact feature mapping required before comparison",
      "endpoint": "Directional-transfer assessment",
      "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun",
      "permanent_baselines": [
        "zero change",
        "train mean",
        "ridge"
      ],
      "primary_metrics": [
        "formula-aligned aggregate directional diagnostics"
      ],
      "eligibility_gate": "Frozen aggregate target effects",
      "current_result": "No directional-transfer support",
      "decision_blocker": "No harmonized cell-population outcome or matched feature space",
      "route": "/resource/benchmarks/external-transfer/"
    },
    {
      "task_id": "G2",
      "label": "Unseen gene family or pathway",
      "state": "REGISTERED_NOT_EXECUTED",
      "split_contract": "held-out related targets with leakage-safe exposure audit",
      "feature_space": "source-defined; exact feature mapping required before comparison",
      "endpoint": "Held-out family response distribution",
      "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun",
      "permanent_baselines": [
        "zero change",
        "train mean",
        "ridge",
        "nearest neighbour"
      ],
      "primary_metrics": [
        "DES",
        "PDS",
        "MAE",
        "calibration",
        "abstention"
      ],
      "eligibility_gate": "Enough related targets to freeze disjoint train and test families",
      "current_result": "No result",
      "decision_blocker": "Current candidate set is too small for a powered family holdout",
      "route": "/resource/benchmarks/generalization/"
    },
    {
      "task_id": "G3",
      "label": "Unseen donor",
      "state": "LOCKED_NO_TRUTH",
      "split_contract": "entire donor withheld",
      "feature_space": "source-defined; exact feature mapping required before comparison",
      "endpoint": "Donor-conditioned response and responder distribution",
      "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun",
      "permanent_baselines": [
        "population mean",
        "donor-agnostic ridge",
        "nearest donor"
      ],
      "primary_metrics": [
        "calibration",
        "donor replication",
        "distribution distance",
        "abstention"
      ],
      "eligibility_gate": "Multiple biological donors with repeated perturbations",
      "current_result": "No result",
      "decision_blocker": "No qualifying donor-resolved DMD perturbation matrix",
      "route": "/resource/benchmarks/generalization/"
    },
    {
      "task_id": "G4",
      "label": "Unseen cell state",
      "state": "LOCKED_NO_TRUTH",
      "split_contract": "train in one myogenic state and test another",
      "feature_space": "source-defined; exact feature mapping required before comparison",
      "endpoint": "State-conditioned population transition",
      "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun",
      "permanent_baselines": [
        "unperturbed trajectory",
        "state mean",
        "additive shift"
      ],
      "primary_metrics": [
        "state occupancy",
        "DES",
        "PDS",
        "MAE",
        "transport distance"
      ],
      "eligibility_gate": "Matched perturbation identities across physical time and state",
      "current_result": "No result",
      "decision_blocker": "Current myogenic trajectory is real but unperturbed",
      "route": "/resource/trajectory/"
    },
    {
      "task_id": "G5",
      "label": "Future batch or external laboratory",
      "state": "LOCKED_NO_TRUTH",
      "split_contract": "prospective time and site shift",
      "feature_space": "source-defined; exact feature mapping required before comparison",
      "endpoint": "Prospective response, uncertainty and abstention",
      "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun",
      "permanent_baselines": [
        "historical mean",
        "ridge",
        "expert selection"
      ],
      "primary_metrics": [
        "calibration",
        "AUPRC",
        "hit enrichment",
        "abstention quality"
      ],
      "eligibility_gate": "Outcome collected after Prediction Card freeze",
      "current_result": "No result",
      "decision_blocker": "No registered prediction or future outcome",
      "route": "/resource/prediction_registry"
    },
    {
      "task_id": "G6",
      "label": "Independent DMD disease line",
      "state": "LOCKED_NO_TRUTH",
      "split_contract": "new disease line or genotype",
      "feature_space": "source-defined; exact feature mapping required before comparison",
      "endpoint": "Molecular, functional, toxicity and replication outcomes",
      "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun",
      "permanent_baselines": [
        "zero",
        "train mean",
        "ridge",
        "expert choice",
        "random selection"
      ],
      "primary_metrics": [
        "functional hit rate",
        "AUPRC",
        "calibration",
        "toxicity miss rate",
        "replication"
      ],
      "eligibility_gate": "Independent DMD candidate perturbations across donors and reagents",
      "current_result": "No result",
      "decision_blocker": "0 independent DMD candidate-perturbation outcomes",
      "route": "/resource/minimum-perturbome/"
    },
    {
      "task_id": "G7",
      "label": "New modality and combination",
      "state": "LOCKED_NO_TRUTH",
      "split_contract": "KO, CRISPRa, drug or double perturbation",
      "feature_space": "source-defined; exact feature mapping required before comparison",
      "endpoint": "Main effects, interaction, synergy class and uncertainty",
      "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun",
      "permanent_baselines": [
        "additive null",
        "pathway mean",
        "nearest neighbour"
      ],
      "primary_metrics": [
        "interaction error",
        "synergy classification",
        "calibration",
        "toxicity"
      ],
      "eligibility_gate": "Observed same-screen singles and doubles or matched genetic–chemical pairs",
      "current_result": "Only an additive no-interaction preview is available",
      "decision_blocker": "No observed candidate doubles or matched DMD drug screens",
      "route": "/resource/perturbation-lab"
    }
  ]
}