Using rocci with your data

Ingestion is duck-typed: rocci accepts what the ecosystem produces without importing any of the producing libraries. That means it works with all the usual scientific computing libraries, but also with custom types - just include an __array__ or .to_numpy() method. Anything array-like works; anything ambiguous raises a RocciError that names the fix.

Containers

You have	It works because
NumPy arrays, Python lists/tuples	used directly
pandas Series / Index	__array__ protocol
polars Series	.to_numpy() duck-call
torch / JAX / CuPy tensors	__dlpack__; CUDA tensors fall back to .detach().cpu().numpy()
xarray / arviz DataArray	__array__ protocol

Column vectors — y.reshape(-1, 1) labels or predict_proba(X)[:, [1]] scores — are recognized and raveled.

Labels (y_true)

Booleans, {0, 1}, and {-1, 1} infer the positive class automatically. Any other pair of values (strings included) works with an explicit pos_label=:

roc_band(diagnosis, score, pos_label="malignant")

More than two distinct labels is an error by design — there is no single ROC curve for a multiclass problem. Use roc_band_ovr for one-vs-rest bands with a family-wise guarantee, or binarize and pass pos_label.

Scores (y_score)

• 1-D scores — used directly. Higher must mean "more positive".
• (n, 2) probability matrices (rows summing to 1, e.g. predict_proba output) — the positive-class column is selected automatically, recorded as an INFO note in band.summary().
• Posterior draws — (draws, n) or (chain, draw, n) arrays (PyMC, arviz) require score_reduce="mean" or "median"; rocci reduces over the draw axes and notes that the band then quantifies sampling uncertainty of the reduced scores, not posterior uncertainty. The Bayesian vignette discusses what that means in practice.
• Multiclass predict_proba ((n, m>2), rows summing to 1) — an error pointing you to roc_band_ovr.

Logits or probabilities?

Either. The envelope band is invariant to strictly monotone transforms of the scores, so logits and their sigmoid give the identical band — no calibration step needed first. (The Working–Hotelling band does not have this property; see Which band should I use?.)

Ties, discreteness, ±inf, NaN

• Ties and discrete scores never fail. The band stays valid and errs conservative; heavy ties (fewer than half the pooled scores distinct) get a TiesWarning saying exactly that. Even a class with entirely constant scores proceeds — the band degenerates to the exact floors and is honestly wide.
• ±inf scores are legal (they sort; a strictly monotone squashing would give the same band anyway).
• NaN raises by default. Pass nan_policy="omit" to drop NaN rows (in either input) with a warning stating the count.

From estimators

from rocci import from_estimator
band = from_estimator(clf, X_test, y_test)

Uses predict_proba when the estimator has it, else decision_function (override with response_method=). Works with scikit-learn pipelines, statsmodels results wrapped in a thin shim, or anything else exposing either method.

Small samples

Classes with fewer than 2 samples are an error; fewer than 20 gets a SmallSampleWarning. The band remains valid at any size — the exact floors take over and the band gets honestly wide instead of quietly overconfident.