One pinned model set.
Three useful platform roles.
Hugging Face source and evidence. Kaggle CPU reproduction. Cloudflare static delivery.
Run, inspect, reproduce
Completed Kaggle notebook · Pinned Kaggle dataset · Cloudflare preview snapshot
7,200 evaluations and 12 negative controls passed on both the M4 and Kaggle CPU. The metric dictionaries match exactly. Cloudflare passed 60 browser checks, including 12 actual download/reimport paths.
Methods, failures and limits · Machine-readable release · Actual Kaggle output · Actual M4 output · Cloudflare browser readback
The original model's large worst-case error remains. This is already-public synthetic data, not new customer validation. The first Kaggle run failed before computation and is retained; version 2 completed. No GPU, private-data publication, signing or GSPC admission.
Keep the roles separate
The Cloudflare site is a static preview on an existing project, not a change to production domains or a paid model server. Kaggle runs the Python calculation without internet; Hugging Face retains the canonical assets and this evidence.
python3 replay_stdlib.py --data /path/to/dataset --out ./results
Inspect the new Python implementation · Notebook source · Fixed input hashes · Release hashes
Do the three platforms still match?
Recorded observation: 2026-11-11T06:05:07.528421+00:00. 85 selected file observations matched the pinned source; the existing Kaggle notebook and its result bytes matched too.
Historical observation, not a live health check. Current status has not been refreshed in this browser.
The Hugging Face head had advanced, while the selected model, data and execution files remained unchanged. The Cloudflare preview is explicitly not the latest full repository release.
Method, rerun command and limits · Recorded verdict · Actual readback · 44 negative and boundary controls
Missing or changed bytes, failed/queued notebooks, moved aliases and expired observations do not become a match. This check does not grant human approval, source authenticity or GSPC admission.
What remains before a real pilot?
The documented research workflows are usable within their stated scope. Pilot approval and production readiness are not established. This is an owner handover, not a new approval system.
At 2026-11-11T06:19:46.870448+00:00, the public human-review metadata still reported individual identity not verified, human presence not established and no execution authority. A configured token, store or signing key does not close that gap.
| Step | Remaining work | Current scoped state |
|---|---|---|
| G0 | Scope and spend boundary | HELD FOR OWNER DECISION |
| G1 | Trusted source, runtime and access binding | PARTIAL RESEARCH ONLY |
| G2 | Authenticated individual and reviewed role | NOT ESTABLISHED |
| G3 | One exact harmless approval-to-effect journey | NOT ESTABLISHED |
| G4 | Task capability and tail-failure qualification | UNDER ACTIVE EVALUATION |
| G5 | Data rights, privacy and storage lifecycle | PARTIAL RESEARCH ONLY |
| G6 | Canonical source, operations and rollback | OWNER ACCEPTANCE REQUIRED |
| G7 | Outside recipient and use-case acceptance | NOT EVIDENCED |
| G8 | Release and optional measurement admission | HELD FOR OWNER DECISION |
Read the complete acceptance runbook · Download the completion handover · Recorded identity metadata · Spend boundary and known existing job · Exact-input dataset audit
Inspect the offline handover checker · 19 document-boundary checks · Source and evidence hashes
The checker validates this document package only; it cannot issue credentials, appoint reviewers, start compute, execute an action or admit measurements. The separate existing billable CPU job is recorded with an unknown exact charge; no new paid compute was started in this pass.