# MEOK: actual Hugging Face, Kaggle and Cloudflare execution

**11 November 2026 · Executed on the existing Apple M4 and Kaggle CPU; Cloudflare static preview delivered and tested.**

## Open the actual outputs

- [Completed Kaggle notebook](https://www.kaggle.com/code/nicktempleman/meok-cross-platform-cpu-replay) — version 2, standard CPU, no GPU/TPU/internet.
- [Public pinned Kaggle dataset](https://www.kaggle.com/datasets/nicktempleman/meok-pinned-process-replay) — unchanged public model/data bytes plus MIT notices.
- [Cloudflare deployment snapshot](https://17c2bf23.csoai-verify.pages.dev) — static preview of the existing public lab, not a new inference service.
- [Canonical HF input snapshot](https://huggingface.co/spaces/csoai/meok-oowm-hybrid-lab/tree/1c2395b61d8d45955b28f05f971e759ddc799b49) — source of all copied research inputs.

## The missing integration addressed

Targeted discovery found prior governance and Academy notebooks, but not a numerical replay of these six public process checkpoints. The Academy work was left untouched. This release adds a separately written Python-standard-library forward pass, input validation, constraints, error aggregation and negative controls. It consumes the published weights/data as bytes; it does not import the existing JavaScript numerical implementation, fit weights or call an LLM provider.

The local M4 run used Python 3.14.7 on arm64. The real Kaggle run used Python 3.13.18 on Linux x86_64. Both executed **7,200 inference evaluations**, validated ten pinned input files, and passed **12 negative controls**. Their resulting metric dictionaries are exactly equal. Against the earlier JS/NumPy references, the largest metric difference is 2.22e-16; the original-bank per-output difference is 4.44e-16. This is numerical reproducibility—not an accuracy improvement or independent organisational certification.

## Results retain the failure distribution

All values below are percentage points of synthetic population share, not chat-answer accuracy.

| Variant | Original mean | Original worst | Later-bank mean | Later-bank worst |
|---|---:|---:|---:|---:|
| original_float64 | 0.287973 | 74.990766 | 0.187381 | 6.177835 |
| refitted_head | 0.274946 | 74.852745 | 0.168533 | 5.260786 |
| shuffled_head_control | 5.152744 | 42.581924 | 5.516802 | 62.860728 |
| weight_float32 | 0.287973 | 74.990769 | 0.187381 | 6.177832 |
| weight_int4 | 4.509438 | 26.617624 | 4.899166 | 38.624623 |
| weight_int8 | 0.329068 | 72.926492 | 0.246134 | 6.983247 |

The roughly 75-percentage-point original-bank tail failure remains visible. Lower average error is not a reason to promote the fitted head. Both banks are already public and generated from the same synthetic process, so this is not a fresh blind test or real-world generalisation result.

## Kaggle execution and failures

The public dataset is ready. The notebook's second version completed and its output files were downloaded and checked against M4 outputs. The published code cells and CPU/no-internet metadata were pulled back and verified. `kaggle-results.json` is the actual remote output, not a local result relabelled as Kaggle execution.

The first submission title collided with the dataset title; it returned HTTP 409 without launching a notebook. After using a distinct title, version 1 failed in Papermill before any model calculation because the notebook lacked a kernelspec. Version 2 adds Python3 kernel metadata; the numerical code and data are unchanged. That failed version and the correction are retained in the evidence. There is no hidden claim that the first run succeeded.

The Kaggle licence label is **Other (specified in description)** because the client did not expose MIT as a supported dataset label. Full upstream MIT LICENSE/NOTICE files are retained. No CC0/public-domain claim or relicensing is made.

## Cloudflare: use the edge for delivery, not an unnecessary model server

The existing `csoai-verify` project now has a **preview branch**, `meok-replay-20261111`. Its production deployment head remains unchanged. No custom domain, production route, private bucket, R2/D1/KV binding, Worker function or Workers AI service was changed or enabled.

The preview serves a pinned copy of the existing HF lab. The source and delivered-file hashes are in `cloudflare-mirror.json`; only the landing HTML gains a clearly labelled mirror notice, with deployment headers and a real 404 page added. Computation remains in the visitor's browser. The preview is a dated snapshot, not an automatic latest-release mirror or a new source of authority.

**Sixty hosted browser checks passed** at desktop and mobile-sized viewports, including **twelve actual downloads** and file-selection/recompute/export/reimport workflows. The same-file, invalid-UTF8, oversized/empty-input, duplicate-key, rehashed-false-metric and delayed-file-read controls passed. Normal tested workflows recorded zero JavaScript errors and zero network write requests. Selected research receipt files were not uploaded by those workflows.

An initial readback assertion compared entire 25-item deployment-list pages; adding a preview displaced the oldest item. The corrected check verifies the unchanged latest production deployment and overlapping ordered history, without claiming an exhaustive history audit. One early direct HTTP request returned 403; later public requests and the real browser workflows succeeded. It is retained as a transient readback failure, not a suppressed successful test.

Cloudflare's official documentation states static asset requests are free; dynamic Worker execution and storage/AI products have separate limits and billing. This release is deliberately static-only. It does not rely on startup credits or enable a paid service. Reference: https://developers.cloudflare.com/pages/functions/pricing/ .

## Reproduce

Download the Kaggle dataset and keep its filenames unchanged. The replay code uses no third-party numerical packages:

```sh
python3 replay_stdlib.py --data /path/to/dataset --out ./results
```

The dataset manifest is pinned to `9ad0d91f9f6a80e89252c897b20dc689b4e6cba0a73babc0a5967c41289b58b2`. Wrong bytes, malformed inputs and false numerical claims are refused. The same code is embedded in the public notebook, whose attached dataset allows it to run without internet. `metrics.csv`, local and remote result JSON, deployment readback and actual browser records are retained here.

## Roles and onward use

Hugging Face remains the versioned source/evidence hub. Kaggle supplies a discoverable executable CPU reproduction and actual remote outputs. Cloudflare supplies a static delivery path for that already-built public interface. These roles complement one another; they do not require merging weights, duplicating the private MEOK runtime or granting a model authority.

Useful later work is to keep the same hash-bound release across distribution surfaces, attach independent user/operator reproduction, and qualify trusted source/context and real human approval-to-effect separately. No automatic synchronization task, training pipeline, paid compute, signing or GSPC admission is created here.