mcaf-ml-ai-delivery

# MCAF: ML/AI Delivery Read root/local `AGENTS.md`, the current ML stage, accepted product outcomes, data constraints, and existing model/pipeline contracts. ## Workflow 1. Separate product, data, and model assumptions; identify the uncertainty or delivery requirement the task must resolve. 2. Use representative data, a documented baseline, meaningful evaluation, reproducible datasets/parameters/artifacts, and leakage-aware splits. 3. Define data quality, harmful-output/uncertainty handling, responsible-AI risks, and stage exit criteria before experiments or implementation. 4. Keep exploratory experiments distinct from production delivery claims. Complete approved data, model, inference, tests, and integration changes before self-review and planned software verification. 5. Review evaluation code and tests for real outcomes, required rejection paths, and model/data contract failures. Obtain approval for material model architecture, serializer, data, dependency, or deployment decisions. 6. Report evidence, limitations, and readiness separately; an offline metric does not prove the production flow. ## References Load only the guidance for the active stage: - Framing: [ML project stages](references/ml-ai-projects.md), [feasibility](references/feasibility-studies.md). - Data: [exploration](references/data-exploration.md). - Experiments: [experiment records](references/model-experimentation.md), [fundamentals](references/ml-fundamentals-checklist.md). - Verification: [data science and MLOps tests](references/testing-data-science-and-mlops-code.md). - Release: [model readiness](references/ml-model-checklist.md), [responsible AI](references/responsible-ai.md).