Learn in a notebook

Download a notebook and its matching support.py into the same directory. The helper is portable and pinned to SDK 0.7; imports do not depend on your checkout. Sources have cleared outputs and default to no-network synthetic transport.

Install and open#

Install the 0.7 release wheel, then add Jupyter to your local environment. This is an optional learning tool, not an SDK runtime dependency. Download support.py and a notebook below. Open it from their shared directory and restart/run all.

python -m pip install jupyterlab
python -m jupyter lab

Notebook	What you learn
Classification	A DataFrame with explicit target, probabilities, shape and reopen.
Regression	Explicit numeric task, finite points, column alignment and pinned reuse.
Async jobs	Await resources, keep durable job ID and reconnect.
Artifacts and usage	Predictor provenance and logical operation accounting, without pricing.
Research datasets	Inspect metadata and download checksum-pinned bytes without private upload.
Native causal discovery	Target-free observations, immutable native graph, separate scores and reconnect.

Execution instructions explain fixture and live modes. Static docs also retain release-local copies in downloads/; the linked release tag remains the canonical notebook source.

Know what ran#

Default fixture mode validates notebook/client plumbing with controlled HTTP transport. It does not execute a foundation model, prove backend security or make scientific/latency claims. Classification uses seed9 and regression seed42, each with 128 training rows, four features and 32 query rows. Native discovery uses the full seed42 128×4 chain/isolate observations; scores and graph semantics remain separate. Research fixture bytes are clearly synthetic and are never called a real Parquet download.

Live mode is an explicit separate opt-in with current task qualification/readiness, locally configured key and HTTPS endpoint. Fits/discoveries retain the created synthetic dataset/predictor/result. No notebook deletes hosted state or stores keys. Clear outputs before saving. Local Ergodic conversion requires the separately approved co-release artifact; absence is shown and never counted as successful causal continuation. Full catalogue, true large batch and five-engine causal packs remain separate gates.

Reproduce checks from source#

CI builds the wheel, installs it, then executes actual disposable Jupyter kernels. It rejects saved outputs and never persists executed predictions.

uv sync --frozen
uv build
uv pip install --python .venv/bin/python --reinstall --no-deps dist/*.whl
.venv/bin/python scripts/check_notebooks.py
.venv/bin/python scripts/build_docs.py

Use the direct environment Python after wheel installation; uv run can restore editable source. The runner forces fixture mode unless --live is supplied; --notebook regression selects one example without counting other notebooks as checked. 0.6 examples remain pinned for users of that release.