- H2O Driverless AI writes an editable AutoDoc report for every experiment: data, validation scheme, feature engineering, final model performance and explanation plots.
- Machine learning interpretability gives reason codes for each prediction, plus Shapley values, surrogate decision trees, partial dependence and disparate impact analysis.
- Enterprise h2oGPTe answers your validators' evidence questions from AutoDocs, validation reports and policies, with a citation to the source page.
- The evidence supports the model risk process you already run; your model risk team still reviews, challenges and approves.
Every model has to answer the same questions
Your validators ask the same things of every model: what data trained it, which features drive the score, whether it was tested for disparate impact, and what changed since the last review.
The answers usually live in notebooks, emails, slide decks and a documentation file written weeks after the model was built. Rebuilding that file by hand is slow, and a validator who cannot find the evidence sends the model back.
How AI model risk documentation works
H2O Driverless AI builds and explains predictive models such as credit, fraud and pricing models, and Enterprise h2oGPTe turns the evidence into cited answers.
Step 1: document while the model trains
AutoDoc writes an editable report per experiment, with feature engineering, validation scores across folds, final model performance, population stability and permutation feature importance. A custom template can follow the layout your model risk policy asks for.
Step 2: explain every prediction
Machine learning interpretability produces K-LIME, Shapley values, variable importance, a surrogate decision tree, partial dependence, ICE and sensitivity analysis. LIME and Shapley reason codes download as CSV, and surrogate tree rules export as text and Python.
Step 3: test for fairness
Disparate Impact Analysis reports the adverse impact ratio against a four-fifths threshold you can adjust, plus error measures by group.
Step 4: make interpretability a setting
The interpretability dial favors simpler features, and monotonicity constraints switch on at higher settings. Compliant mode restricts the search to GLM or RuleFit models.
Step 5: track the model in production
H2O MLOps keeps lineage, model monitoring for model drift and an audit trail.
Step 6: answer evidence questions
Load AutoDocs, validation reports, monitoring reports and model policies into a private collection, and Enterprise h2oGPTe answers a validator's question with the source passage cited.
Explaining a model built elsewhere
Machine learning interpretability can also explain a model built outside H2O, given its target and prediction columns. That lets you add reason codes to a champion model while you test a challenger, without rebuilding it in a new tool first.
Where the evidence for each question lives
A validator's question about what data trained the model points to the AutoDoc data summary and validation scores; a question about why one customer got a particular score points to the LIME and Shapley reason codes for that row; and a question about whether the model treats groups differently points to the Disparate Impact Analysis.