Executed notebook examples

Five source-candidate examples demonstrate task-gated workflows:

• Classification: synthetic table upload, durable fit, probabilities, repeated prediction and reopen.
• Regression: separately qualified task, finite mean points, named-column alignment and pinned reopen.
• Async resources and jobs: async requests, durable identity and reconnect using SDK0.2.

• Artifacts and usage: fitted provenance and logical operation accounting without pricing or deletion.

• Causal discovery: target-free upload, immutable native graphs, explicit owned score access, provenance and reconnect. Live ArrowFM execution and optional local Ergodic continuation remain separate gates.

Execution instructions cover default controlled transport, explicit bounded live mode and credential-safe source handling. CI executes real Jupyter kernels from an installed SDK wheel using nbclient, keeps outputs in memory/temp directories and rejects saved source outputs. Synthetic fixture execution proves example/client plumbing, not FM quality or backend security. Separate live evidence is needed for real model behavior.

Further model swapping, true batch workflows and five-engine causal-to-ergodic examples are later gates requiring their actual supported contracts and runtimes. No placeholder notebook is counted as working or silently skipped to close the complete catalogue milestone. Notebook sources have cleared execution counts and outputs. Download/open notebooks from the matching reviewed SDK revision for reuse.