Model Evaluation and Promotion Pipeline

# ML Model Evaluation Workflow
# Evaluate a candidate model against baseline and decide promotion or rollback.
osop_version: "2.0"
id: model-evaluation
name: Model Evaluation and Promotion Pipeline

nodes:
  - id: load_test_data
    type: db
    purpose: Load held-out test dataset from the data warehouse
    runtime:
      engine: bigquery
      connection: project.dataset.test_set
    outputs:
      - test_dataset

  - id: run_inference
    type: cli
    purpose: Run candidate model inference on the test dataset
    runtime:
      command: "python inference.py --model ${candidate_model} --data ${test_dataset}"
    inputs:
      - test_dataset
      - candidate_model
    outputs:
      - predictions
    timeout_sec: 300

  - id: compute_metrics
    type: cli
    purpose: Compute accuracy, precision, recall, F1, and latency metrics
    runtime:
      command: "python metrics.py --predictions ${predictions} --ground-truth ${test_dataset}"
    inputs:
      - predictions
      - test_dataset
    outputs:
      - metric_results

  - id: load_baseline
    type: db
    purpose: Fetch baseline metrics from the model registry
    runtime:
      engine: postgres
      connection: postgresql://registry:5432/models
    outputs:
      - baseline_metrics

  - id: compare_baseline
    type: agent
    purpose: Compare candidate metrics against baseline and produce a verdict
    runtime:
      provider: anthropic
      model: claude-sonnet-4-20250514
    inputs:
      - metric_results
      - baseline_metrics
    outputs:
      - comparison_verdict
    explain: "Agent interprets metrics holistically, not just single-threshold."

  - id: generate_report
    type: agent
    purpose: Generate a human-readable evaluation report with charts and tables
    runtime:
      provider: anthropic
      model: claude-sonnet-4-20250514
    inputs:
      - metric_results
      - comparison_verdict
    outputs:
      - eval_report

  - id: promote_model
    type: system
    purpose: Promote candidate model to production in the model registry
    runtime:
      tool: model-registry-cli
    inputs:
      - candidate_model
    outputs:
      - promotion_status

  - id: rollback_alert
    type: system
    purpose: Send alert that candidate model did not pass evaluation
    runtime:
      tool: pagerduty
    inputs:
      - eval_report
    outputs:
      - alert_status

edges:
  - from: load_test_data
    to: run_inference
    mode: sequential

  - from: run_inference
    to: compute_metrics
    mode: sequential

  - from: load_baseline
    to: compare_baseline
    mode: sequential
    explain: "Baseline loads in parallel with inference; both feed into compare."

  - from: compute_metrics
    to: compare_baseline
    mode: sequential

  - from: compare_baseline
    to: generate_report
    mode: sequential

  - from: generate_report
    to: promote_model
    mode: conditional
    condition: "comparison_verdict.passed == true"

  - from: generate_report
    to: rollback_alert
    mode: conditional
    condition: "comparison_verdict.passed == false"
    explain: "Route to promotion or rollback based on evaluation outcome."