SDK guide

Client() defaults to https://welt.ergodic.dev; explicit base_url overrides WELT_BASE_URL. Authenticated methods use WELT_API_KEY or an explicit key. Always close it or use a context manager. Client.models() reports actual capabilities, not blanket task support.

from welt import Client

with Client() as client:
    data = client.upload(columns=columns, rows=rows, target=target_name)
    job = client.submit_fit(data['id'], model='tabicl-v2', task='classification', seed=9)
    predictor = job.result(timeout=600)
    metadata = client.predictor(predictor['id'])

Classifier and Regressor support fit/predict and predict_details; Classifier supports predict_proba. Labels/probabilities follow persisted class order. DataFrame prediction columns may be safely reordered; missing/extra names or width/type changes reject. Arrays follow positional order. CSV/Parquet paths are convenient in-memory reads today, not resumable streaming upload. submit_fit(X, y) or submit_fit(train, target="label") returns a durable Job without blocking; it owns an HTTP client, so close job.client in a finally block when finished. Blocking estimator fit closes that client internally. Explicit Client uploads permit dataset-ID reuse.

Inspect job.id/job.inspect(), reconnect with client.job(id) and retrieve with result(). A polling deadline is local; it does not cancel durable server execution. job.cancel() explicitly requests cancellation. Typed errors distinguish invalid input, unavailable model, authentication, capacity and execution failure; IDs allow operator diagnosis without exposing data values. PredictionPendingError carries a job_id for reconnect after server-side small-prediction timeout.

Predictors retain exact model/config/seed identity. Classifier.from_predictor(id) reopens a persisted fit. A new fit creates a new predictor; cache eviction does not mean deletion. Current SDK resources include datasets(), predictors(), jobs() and usage(); metadata inspection/export_metadata() returns authorized JSON-compatible server provenance only, never weights/context/credentials. Explicit owned dataset/causal-result deletion is available; recipe/deletion-plan transport remains separate.

Only controlled tests establish current clone/Pipeline/CV and construction behavior. The full official estimator/task/version matrix, batch/resumable files, remaining causal engines and public Ergodic continuation remain gates. AsyncClient/AsyncJob and native ArrowFM/CDFM/AVICI results retain their independently qualified narrow boundaries; the supported notebook checks remain distinct from real model qualification. SDK 0.7 retains the research catalogue without including unreleased CSV0.4 or batch0.5. No conformal uncertainty or full catalogue availability is claimed by this SDK release.

Safe metadata GET requests have bounded automatic retries (default two, maximum five), respecting Retry-After up to the configured wait bound; a larger hint returns an actionable typed error rather than retrying early. POST mutations never auto-retry. Reusing a Client submission requires explicit idempotency_key preservation by the caller; omitted keys are freshly generated per call. Estimator fit/submit_fit makes a new dataset upload and logical fit, so call Client.submit_fit on the existing dataset for an intentional request replay.

The capability boundary, onboarding guide, executed notebook examples and generated installed-package reference explain supported behavior. The source examples default to synthetic transport.