shadow-ds-govern

# Shadow DS Govern Model-risk council for data-science pipelines. Deterministic. No LLM. ## When to use The user's request contains an ML artifact reference — at minimum a model + feature list — and asks any variant of: - "Should we ship this?" - "Is this defensible for production?" - "Would compliance sign off on this?" - "Are we above the disparate-impact ratio?" ## Required input An `MLArtifactRef` object. Minimum viable payload: ```json { "artifact_id": "mlflow-run-abc123", "model_type": "xgboost.XGBClassifier", "task": "credit_scoring", "feature_columns": ["fico", "dti", "ltv", "amount", "sector"] } ``` Optional but recommended (missing → REWORK on the relevant voice): - `drift_snapshot.psi` — Population Stability Index (Data Steward) - `calibration.brier` — out-of-sample Brier score (Model Validator) - `disparate_impact.aim_ratio` — adverse-impact ratio (Fair-ML Auditor) - `ops_metrics.p95_ms` — inference latency (Ops Realist) ## What it does Five voices, each applying a declarative scorer: | Voice | Trigger | Verdict | |---|---|---| | Data Steward | PSI > 0.25 | REWORK | | Model Validator | Brier > 0.25 | REWORK | | Fair-ML Auditor | adverse-impact ratio < 0.80 | **BLOCK** (EEOC 80% rule) | | Reproducibility Critic | artifact_id or feature_columns missing | REWORK | | Ops Realist | p95 > 1000 ms | REWORK | Resolver: ANY BLOCK → BLOCK; ANY REWORK → REWORK; ALL SHIP → SHIP. Same conservatism as banking-side block > escalate > approve. ## The named i...