Evidence Benchmark
MAMUT includes a lightweight benchmark script for release diagnostics. The goal is not to claim state-of-the-art AutoML performance. The goal is to verify that the selected model beats trivial baselines, that stronger baselines are visible when they challenge the selection, and that score stability is reported with descriptive resampling intervals.
Run the benchmark from the repository root:
uv run python scripts/benchmark_evidence.py --format markdown
The default run uses:
sklearn
breast_cancer,digits, andwinedatasetsfixed
random_state=42balanced_accuracyas the selection metricholdout_size=0.2for final evaluationfinal refit on all non-holdout modeling rows before holdout scoring
one random-search iteration for speed
three-fold repeated stratified CV with one repeat for score stability
the lightweight
quickcandidate profile (logistic regression, random forest, extra trees, and Gaussian naive Bayes)
Example Diagnostic Output
The following output was generated from the locked development environment for a release validation pass:
| dataset | samples | features | classes | selected_model | holdout_score | best_baseline | best_baseline_score | repeated_cv_mean | repeated_cv_ci | guidance | leakage_warnings |
| ------------- | ------- | -------- | ------- | ------------------ | ------------- | ------------------- | ------------------- | ---------------- | -------------- | ---------- | ---------------- |
| breast_cancer | 569 | 30 | 2 | RandomForestClassifier | 0.943 | Logistic Regression | 0.953 | 0.958 | [0.879, 1.000] | challenged | 0 |
| digits | 1797 | 64 | 10 | RandomForestClassifier | 0.969 | Random Forest | 0.969 | 0.962 | [0.939, 0.986] | challenged | 0 |
| wine | 178 | 13 | 3 | LogisticRegression | 1.000 | Logistic Regression | 1.000 | 0.981 | [0.934, 1.000] | confirmed | 0 |
Interpretation
confirmed means no evidence baseline exceeded the selected model by the
configured practical margin. challenged means a baseline matched or beat the
selected model strongly enough to require review. A challenge is useful signal:
it prevents MAMUT from presenting a validation-selected model as stronger than
the evidence supports.
On breast_cancer, a simple logistic-regression baseline beats the selected
random-forest candidate on the holdout split, and MAMUT surfaces that
challenge. On digits, a random-forest baseline matches the selected
candidate closely enough to require review. On wine, the holdout score is
saturated while score-stability evidence remains visible. This is the intended
behavior: attractive holdout results do not suppress baseline or stability
checks.