Requirements are canonicalised against an approved offline package catalogue and stored with Python/runtime profile and SHA-256 identity. A failed candidate does not replace the previous ready environment.
ANPy governed Python
Controlled Python work in a project-scoped runtime.
ANPy gives Python analysis a known environment, notebook revision, kernel lifecycle and CellRun record while keeping production isolation and package policy under platform control.
Canonical nbformat 4 documents use explicit manual save and strong ETags for concurrent writes. Checkpoints and restores create immutable versions, while the saved notebook head remains an explicit mutable revision. A conflict preserves the local draft and requires a chosen reload or reapply path.
The trusted runtime-manager validates workspace, notebook and environment identity before Start, Interrupt, Restart or Stop. Browser input cannot select host paths, containers or control credentials.
After a new kernel, bounded replay can reconstruct prior successful, unchanged setup cells only when revision, source hash, environment and run evidence agree. Stale or incomplete proof fails closed.
Each submitted CODE cell retains source identity, notebook revision, environment and kernel context, bounded stdout/stderr metadata, duration and terminal state. The current browser has no live stdout stream or complete DataFrame/chart renderer.
Environment is part of the analysis identity.
Requirements are canonicalised against an approved offline package catalogue and stored with Python/runtime profile and SHA-256 identity. A failed candidate does not replace the previous ready environment.
Requirements may be pasted or uploaded into the environment contract.
Package resolution is fixed and offline against the approved catalogue; arbitrary internet installation is not allowed.
Workspace, Python version, normalised requirements and runtime profile produce a stable environment identity.
Notebook state is revisioned, not silently overwritten.
Canonical nbformat 4 documents use explicit manual save and strong ETags for concurrent writes. Checkpoints and restores create immutable versions, while the saved notebook head remains an explicit mutable revision. A conflict preserves the local draft and requires a chosen reload or reapply path.
Kernel lifecycle is controlled outside the browser.
The trusted runtime-manager validates workspace, notebook and environment identity before Start, Interrupt, Restart or Stop. Browser input cannot select host paths, containers or control credentials.
Safe replay is not a memory snapshot.
After a new kernel, bounded replay can reconstruct prior successful, unchanged setup cells only when revision, source hash, environment and run evidence agree. Stale or incomplete proof fails closed.
Stop and Restart do not persist Python process memory.
Eligible unchanged predecessor cells may replay after a new kernel generation.
Changed source, environment or incomplete predecessor evidence blocks replay instead of guessing namespace state.
CellRun keeps code execution bounded and inspectable.
Each submitted CODE cell retains source identity, notebook revision, environment and kernel context, bounded stdout/stderr metadata, duration and terminal state. The current browser has no live stdout stream or complete DataFrame/chart renderer.
ANPy keeps its present boundary visible.
It is not a claim of full MLOps, GPU scheduling, model serving, arbitrary internet package installation, implicit Drive mounting or complete shared-notebook collaboration. Selecting a Drive item is a transient current-item hint, not pinned data lineage or a kernel mount; canonical documents are owner-scoped while surrounding runtime paths do not yet have uniform owner attribution.
Next step
Start with the flow that cannot afford ambiguity.
Bring the systems, owners, rules, delivery obligations and evidence requirements that matter.
Discuss a critical flow