Medtech Model Evidence Export

skill_manifest.yaml

# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

id: medagent.medtech_model_evidence_export
version: 0.2.0
upstream_refs:
  - kind: pypi_package
    name: mlflow
    version_constraint: ">=2.10,<4"
license: Apache-2.0
intended_use:
  summary: >
    Post-hoc export of sanitized Medical AI inference-run metadata, evaluation
    evidence, reproducibility parameters, quality metrics, and explicitly
    approved review artifacts through the documented MLflow Python API.
  scope: development
  not_for:
    - live training-curve tracking
    - unapproved medical-image upload
    - model registration
    - clinical interpretation
    - regulatory submission
inputs:
  - name: evidence_source
    type: path
    description: Evidence pack, trusted-run root containing skill_run/, or direct result JSON.
  - name: mode
    type: string
    description: dry-run, local, or databricks.
    optional: true
    default: dry-run
  - name: artifact_policy
    type: string
    description: metadata, preview, or all; live preview/all requires explicit confirmation.
    optional: true
    default: metadata
outputs:
  - name: export_result
    type: json
    schema: validators/output_schema.json
runtime:
  language: python
  python: ">=3.10"
  entrypoint: scripts/export_evidence_pack.py
  args:
    - "${python}"
    - "${script}"
    - "${fixture}"
    - "--mode"
    - "dry-run"
    - "--artifact-policy"
    - "metadata"
  dependencies:
    mlflow: ">=2.10,<4"
    nibabel: ">=4,<6"
    numpy: ">=1.24,<3"
  test_dependencies:
    jsonschema: ">=4"
    pytest: ">=7"
  env_optional:
    - DATABRICKS_HOST
    - DATABRICKS_TOKEN
    - MLFLOW_TRACKING_URI
  side_effects:
    pip_packages:
      - "mlflow>=2.10,<4"
      - "nibabel>=4,<6"
      - "numpy>=1.24,<3"
    local_writes:
      - {path: "<current-working-directory>/mlruns", approx_mb_max: 1024, optional: true}
    home_writes: []
    network_endpoints:
      - "https://<caller-provided-mlflow-or-databricks-workspace>"
    requires_docker: false
    requires_gpu: none
    environment:
      clean_environment_required: false
      clean_environment_recommended: false
      modifies_active_python_environment: false
      user_environment_modification_ok: false
      recommended_isolation: caller-selected Python environment
limitations:
  - The exporter mirrors existing inference evidence; it does not evaluate model quality.
  - Preview and raw medical artifacts require an explicit policy and live-upload confirmation.
  - NIfTI statistics are sampled for bounded memory use and are engineering heuristics only.
  - Databricks authentication, retention, and experiment permissions belong to the caller.
  - The committed dry-run fixture does not validate a live MLflow or Databricks service.
validation:
  expected_runtime_seconds:
    min: 0.0
    max: 15.0
  sanity_checks:
    - {path: skill, eq: medtech_model_evidence_export}
    - {path: status, eq: dry_run}
    - {path: mode, eq: dry-run}
    - {path: source.pack_kind, eq: skill_run}
    - {path: source.skill_id, eq: medagent.nv_generate_ct_rflow}
    - {path: params, length_gte: 5}
    - {path: metrics, length_gte: 5}
    - {path: artifact_plan.policy, eq: metadata}
    - {path: artifact_plan.metadata_count, gte: 5}
    - {path: note, matches: "Post-hoc MLflow evidence export"}
    - {path: mlflow.available, exists: true}
  expected_cost:
    wall_seconds: {max: 15}
    cpu_seconds: {max: 15}
    rss_mb_peak: {max: 500}
    gpu_seconds: {max: 0}
    gpu_memory_mb_peak: {max: 0}
  env_pin:
    nibabel: ">=4,<6"
    numpy: ">=1.24,<3"
  reproducibility:
    mode: repeat
    fixture: fixtures/sample_pack
    runs: 2
    reason: Dry-run metadata extraction and export planning are deterministic for a fixed pack.