Regression and task-specific availability

The SDK uses the same dataset/preparation/predictor lifecycle for classification and regression. A model family name alone does not imply both tasks are ready. Read its task_profiles entry: regression must be available, have a matching ready worker and report its exact version and limits before preparing a predictor. The SDK never converts regression into classification or silently selects another model.

The bounded TabICLv2 regression recipe returns one finite mean point per query row in target units. Classification classes/probabilities are absent. The underlying model's native quantile channels are not offered as calibrated intervals or conformal coverage by this point-prediction interface. Each task has its own checkpoint, version, qualification and fitted context; a classifier predictor cannot be reused as a regressor.

Use Regressor(model="tabicl-v2") with normal fit(X, y)/predict(X) when the service advertises the qualified task. It uses synchronous sklearn conventions. Existing estimators remain pinned; a new task, seed, configuration or model needs a new fit. Array inputs use positional schema; named DataFrame columns must match the fitted names and are safely reordered. Invalid, missing or extra features reject instead of being silently dropped.

The regression notebook uses the frozen synthetic planted fixture: seed42, 128 training rows, four numeric features and 32 query rows. It checks finite point outputs, null class/probability fields, named-column reordering, pinned version and reopen agreement. It imposes no scientific accuracy threshold and does not force a worker restart. Default CI uses a synthetic HTTP fixture; live opt-in requires the separately qualified real regressor. Accepted synthetic server state remains retained.

Other selected families, task swapping across families, larger envelopes and native uncertainty outputs require their own exact artifact/runtime qualification. This example does not complete the entire model catalogue or regression release matrix.