{ "schema": "nmd-vcell-model-evaluation-contract/1.0", "source_schema": "nmd-vcell-model-evaluation-contract-source/1.0", "contract_id": "MODEL-EVALUATION-CONTRACT:RESEARCH-OS:V201", "updated": "2026-09-15", "local_candidate": { "state_id": "LOCAL:RESEARCH_OS:V201", "deployment_state": "LOCAL_CANDIDATE_NOT_PRODUCTION" }, "authority_rule": "This contract resolves model admission, metric applicability and generalization coverage from existing released registries. It does not run a model, rescore a benchmark or create a DMD prediction.", "adapter_registry": { "adapter_count": 5, "execution_authorized_count": 0, "records": [ { "adapter_id": "ADAPTER:TXPERT:V1", "model": "TxPert", "interface_state": "HISTORICAL_RUN_ONLY", "gateway_decision": "REJECT_UNTIL_NEW_PREREGISTRATION_OR_UNTOUCHED_TRUTH", "compatible_task_ids": [ "G0" ], "execution_state": "HISTORICAL_LOCAL_GATE_FAIL_NO_REPEAT", "released_decision": "STOP_UNTIL_NEW_PREREGISTRATION_OR_UNTOUCHED_TRUTH", "claim_ceiling": "E0_EXECUTION_ONLY", "receipt_count": 0, "source_state": "HISTORICAL_LOCAL_GATE_FAIL_NO_RETRY", "execution_authorized": false }, { "adapter_id": "ADAPTER:STATE:V1", "model": "STATE", "interface_state": "HISTORICAL_RECEIPTS_ONLY", "gateway_decision": "REJECT_NO_ADVANCEMENT", "compatible_task_ids": [ "G0" ], "execution_state": "HISTORICAL_RECEIPTS_MIGRATED", "released_decision": "NO_ADVANCEMENT", "claim_ceiling": "E0_EXECUTION_ONLY", "receipt_count": 8, "source_state": "HISTORICAL_RUNS_MIGRATED", "execution_authorized": false }, { "adapter_id": "ADAPTER:STACK:V1", "model": "Stack", "interface_state": "FORWARD_PATH_SMOKE_ONLY", "gateway_decision": "REJECT_NO_TASK_COMPATIBLE_CHECKPOINT", "compatible_task_ids": [], "execution_state": "LOCAL_NVME_CORE_BASE_FORWARD_PASS_NO_CHECKPOINT", "released_decision": "NO_NMD_B1_BENCHMARK_NO_COMPATIBLE_PRETRAINED_GENETIC_CHECKPOINT", "claim_ceiling": "E0_EXECUTION_ONLY", "receipt_count": 2, "source_state": "SOURCE_PINNED_LOCAL_NVME_CORE_PASS", "execution_authorized": false }, { "adapter_id": "ADAPTER:TAHOE-X1:V1", "model": "Tahoe-x1", "interface_state": "INPUT_CONTRACT_FAILED", "gateway_decision": "REJECT_RAW_COUNTS_REQUIRED", "compatible_task_ids": [], "execution_state": "SOURCE_PINNED_INPUT_CONTRACT_FAIL", "released_decision": "DO_NOT_RUN_ON_LOG_NORMALIZED_INPUT; ACQUIRE_RAW_COUNTS_FIRST", "claim_ceiling": "E0_SOURCE_ONLY", "receipt_count": 2, "source_state": "SOURCE_PINNED_INPUT_CONTRACT_AUDIT", "execution_authorized": false }, { "adapter_id": "ADAPTER:SLIM:V1", "model": "SLIM", "interface_state": "PRIMARY_SOURCE_PENDING", "gateway_decision": "REJECT_SOURCE_NOT_VERIFIED", "compatible_task_ids": [], "execution_state": "PRIMARY_SOURCE_PENDING", "released_decision": "NOT_ELIGIBLE_FOR_EXECUTION", "claim_ceiling": "E0_SOURCE_ONLY", "receipt_count": 0, "source_state": "PRIMARY_SOURCE_PENDING", "execution_authorized": false } ] }, "gateway": { "gateway_id": "MODEL-GATEWAY:NMDVCELL:V1", "mode": "CONTRACT_ONLY_NO_HOSTED_INFERENCE", "default_decision": "REJECT", "execution_endpoint_available": false, "dmd_prediction_route_open": false, "required_gates": [ "registered task contract", "source revision and licence identity", "exact input contract pass", "task-compatible weights or checkpoint", "immutable split and feature mapping", "permanent baseline comparison", "metric applicability and independent-unit contract", "receipt, uncertainty and claim-ceiling record" ], "unregistered_request_policy": "REJECT_NO_REGISTERED_TASK_AND_MODEL_RUN", "current_dmd_decision": "ABSTAIN_NO_CALIBRATED_DMD_MODEL_RUN", "released_model_run_count": 6, "calibrated_dmd_model_run_count": 0, "matched_dmd_candidate_response_truth": "MISSING" }, "metric_lab": { "comparison_rule": "A metric is comparable only within the same task, split, feature space, endpoint and independent-unit contract. Missing metrics stay locked and no global score is computed.", "global_score_permitted": false, "families": [ { "family_id": "AGGREGATE_ERROR", "label": "Aggregate error", "metric_ids": [ "RMSE", "MAE" ], "state": "EXECUTED_BOUNDED", "eligible_task_ids": [ "G0", "G1" ], "boundary": "Aggregate response vectors only; lower error does not establish direction, disease transfer or function." }, { "family_id": "RESPONSE_DIRECTION", "label": "Response direction", "metric_ids": [ "raw cosine", "residual cosine", "delta Pearson", "delta Spearman" ], "state": "EXECUTED_BOUNDED", "eligible_task_ids": [ "G0", "G1" ], "boundary": "Same-context and external aggregate diagnostics remain separate and do not establish DMD transport." }, { "family_id": "PERTURBATION_RETRIEVAL", "label": "Perturbation retrieval", "metric_ids": [ "PDS L1 raw" ], "state": "DIAGNOSTIC_SENSITIVITY_ONLY", "eligible_task_ids": [ "G1" ], "boundary": "Formula-aligned aggregate diagnostic; not an official cell-eval or Virtual Cell Challenge score." }, { "family_id": "SAMPLING_UNCERTAINTY", "label": "Sampling uncertainty", "metric_ids": [ "paired bootstrap interval", "paired sign-flip test", "target bootstrap interval" ], "state": "EXECUTED_BOUNDED", "eligible_task_ids": [ "G0", "G1" ], "boundary": "Intervals quantify the registered benchmark unit only, not donor, disease-stage or clinical uncertainty." }, { "family_id": "CELL_LEVEL_DE_DISTRIBUTION", "label": "Cell-level and DE distribution", "metric_ids": [ "DES", "DEG AUPRC", "single-cell distribution distance", "cell-state proportion recovery" ], "state": "LOCKED_CELL_LEVEL_OR_DE_TRUTH_MISSING", "eligible_task_ids": [], "boundary": "Aggregate predictions cannot be relabelled as cell-level or differential-expression truth." }, { "family_id": "CALIBRATION_ABSTENTION", "label": "Calibration and abstention", "metric_ids": [ "calibration", "coverage", "selective risk", "abstention quality" ], "state": "PARTIAL_DIAGNOSTIC_NOT_DMD_CALIBRATION", "eligible_task_ids": [ "G1" ], "boundary": "Historical selective-risk diagnostics do not constitute a calibrated DMD predictor." }, { "family_id": "FUNCTIONAL_OUTCOME", "label": "Functional and safety outcome", "metric_ids": [ "functional hit rate", "toxicity miss rate", "replication", "synergy classification" ], "state": "LOCKED_MATCHED_OUTCOME_MISSING", "eligible_task_ids": [], "boundary": "No directly measured candidate-conditioned DMD functional outcome exists." } ] }, "generalization_cube": { "cube_id": "GENERALIZATION-CUBE:NMDVCELL:G0-G7:V1", "representation": "SPARSE_REGISTERED_TASK_CELLS_NOT_CARTESIAN_SCORE", "axes": [ "perturbation novelty", "biological context", "donor or line", "dataset or laboratory", "disease background", "modality or combination" ], "cells": [ { "task_id": "G0", "primary_axis": "perturbation novelty", "coordinate": "SAME_CONTEXT_HELD_OUT_TARGET", "label": "Same-context target holdout", "state": "EXECUTED_LIMITED", "ladder_state": "DEBUG_ONLY_NOT_PRIMARY", "split_contract": "repeated-fold-v2.2 target holdout", "endpoint": "Held-out 2,000-feature mean response vector", "independent_unit": "held-out perturbation target; cells do not increase the target denominator", "primary_metrics": [ "RMSE", "raw cosine", "residual cosine", "strict 16-outcome gate" ], "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", "primary_axis": "dataset or laboratory", "coordinate": "EXTERNAL_UNSEEN_PERTURBATION", "label": "External unseen perturbation", "state": "EXECUTED_PARTIAL_UNSUPPORTED", "ladder_state": "IMPLEMENTED_HEPG2", "split_contract": "new perturbation in an external aggregate object", "endpoint": "Directional-transfer assessment", "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun", "primary_metrics": [ "formula-aligned aggregate directional diagnostics" ], "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", "primary_axis": "perturbation novelty", "coordinate": "UNSEEN_GENE_FAMILY_OR_PATHWAY", "label": "Unseen gene family or pathway", "state": "REGISTERED_NOT_EXECUTED", "ladder_state": "REGISTERED_SPLIT_DATA_NOT_EXECUTED", "split_contract": "held-out related targets with leakage-safe exposure audit", "endpoint": "Held-out family response distribution", "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun", "primary_metrics": [ "DES", "PDS", "MAE", "calibration", "abstention" ], "current_result": "No result", "decision_blocker": "Current candidate set is too small for a powered family holdout", "route": "/resource/benchmarks/generalization/" }, { "task_id": "G3", "primary_axis": "donor or line", "coordinate": "UNSEEN_DONOR", "label": "Unseen donor", "state": "LOCKED_NO_TRUTH", "ladder_state": "REGISTERED_DONOR_RESOLVED_TRUTH_MISSING", "split_contract": "entire donor withheld", "endpoint": "Donor-conditioned response and responder distribution", "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun", "primary_metrics": [ "calibration", "donor replication", "distribution distance", "abstention" ], "current_result": "No result", "decision_blocker": "No qualifying donor-resolved DMD perturbation matrix", "route": "/resource/benchmarks/generalization/" }, { "task_id": "G4", "primary_axis": "biological context", "coordinate": "UNSEEN_CELL_STATE", "label": "Unseen cell state", "state": "LOCKED_NO_TRUTH", "ladder_state": "REGISTERED_MATCHED_TRUTH_MISSING", "split_contract": "train in one myogenic state and test another", "endpoint": "State-conditioned population transition", "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun", "primary_metrics": [ "state occupancy", "DES", "PDS", "MAE", "transport distance" ], "current_result": "No result", "decision_blocker": "Current myogenic trajectory is real but unperturbed", "route": "/resource/trajectory/" }, { "task_id": "G5", "primary_axis": "dataset or laboratory", "coordinate": "FUTURE_BATCH_OR_EXTERNAL_LAB", "label": "Future batch or external laboratory", "state": "LOCKED_NO_TRUTH", "ladder_state": "REGISTERED_HARMONIZED_TRUTH_MISSING", "split_contract": "prospective time and site shift", "endpoint": "Prospective response, uncertainty and abstention", "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun", "primary_metrics": [ "calibration", "AUPRC", "hit enrichment", "abstention quality" ], "current_result": "No result", "decision_blocker": "No registered prediction or future outcome", "route": "/resource/prediction_registry" }, { "task_id": "G6", "primary_axis": "disease background", "coordinate": "INDEPENDENT_DMD_DISEASE_LINE", "label": "Independent DMD disease line", "state": "LOCKED_NO_TRUTH", "ladder_state": "REGISTERED_DMD_PERTURBATION_TRUTH_MISSING", "split_contract": "new disease line or genotype", "endpoint": "Molecular, functional, toxicity and replication outcomes", "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun", "primary_metrics": [ "functional hit rate", "AUPRC", "calibration", "toxicity miss rate", "replication" ], "current_result": "No result", "decision_blocker": "0 independent DMD candidate-perturbation outcomes", "route": "/resource/minimum-perturbome/" }, { "task_id": "G7", "primary_axis": "modality or combination", "coordinate": "NEW_MODALITY_OR_COMBINATION", "label": "New modality and combination", "state": "LOCKED_NO_TRUTH", "ladder_state": "REGISTERED_OBSERVED_DOUBLE_TRUTH_MISSING", "split_contract": "KO, CRISPRa, drug or double perturbation", "endpoint": "Main effects, interaction, synergy class and uncertainty", "independent_unit": "study/dataset-defined biological unit; must be recorded by each ModelRun", "primary_metrics": [ "interaction error", "synergy classification", "calibration", "toxicity" ], "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" } ], "scalar_collapse_permitted": false, "counts": { "total": 8, "executed": 2, "registered_not_executed": 1, "locked_no_truth": 5 } }, "source_objects": { "adapter_status": { "path": "/resource/api/v1.1/model_execution_status_registry.json", "sha256": "0e2d0e25be4e421eca5d8b411c8e1292e745e1736d594699ef2870c8b160433e" }, "model_runs": { "path": "/resource/api/v1.1/model_run_registry.json", "sha256": "a429808f22b477ff67b1f5b84ef5c5dce082401e5c44f79fdb2fabda6539d850" }, "task_contracts": { "path": "/resource/api/v1.1/benchmark_task_registry_v2.json", "sha256": "1b831ac73203486fc239f52f7a86a5765d8a88380d58a43052bafa95e00e2917" }, "generalization_lineage": { "path": "/resource/api/v1.1/generalization_ladder.json", "sha256": "7a4749e145dfec23b21b77455bbf9c98f0dc8cb047793fd7e911093f511d9181" }, "metric_diagnostic": { "path": "/resource/api/v1.1/vcc_metric_diagnostic.json", "sha256": "f4ba29fa92c945ed99bf30f50b7775a613863c416104ac4dd94c2afa91a6a944" }, "route_execution": { "path": "/resource/api/v1.1/model_route_execution_v01.json", "sha256": "3c4b8f393649e81d674dd50b964e7a09119f16e4b5d24d768f63baccbca76ac7" } }, "claim_boundary": "The Model Evaluation Gateway reports admission, metric and truth availability from released objects. It creates no inference endpoint, new benchmark result, DMD prediction, experiment authorization or biological claim.", "does_not_imply": [ "a runnable hosted model", "a newly eligible adapter", "a global model score", "DMD generalization", "clinical or therapeutic validity" ], "new_model_run_count": 0, "new_scientific_claim_count": 0, "model_evaluation_contract_sha256": "05934cbd7df873b91cfe6290e82e65cf2226a438abf16049a7565a8bcd24745d" }