# MEOK evidence and precision research v0.3

Open [index.html](index.html) for the browser replay and eight-question practice pack. [WHITEPAPER.md](WHITEPAPER.md) explains methods, results and limits. The original hybrid lab remains one directory above; its source and production defaults are not replaced.

## What actually ran
A small output-layer fit on the existing authorized M2 CPU; six checkpoint variants on 560 original and 640 fresh synthetic transitions; 36 configuration/cache/training/quorum/legacy-boundary checks; and 38 independent Node checks covering 7,200 inference evaluations. The fitted head changes 68 parameters and freezes 912. It is retained, not promoted. No new Hugging Face job, paid hardware, Cosmos inference or model weight merge occurred.

The HF-hosted page replays inference on the visitor's CPU. It does not claim that training occurred on an HF GPU or that the static page is an MCP server. The existing private MCP workspace is preserved; this new candidate has not been silently installed there.

## Reproduce without a remote job
Requires Python with NumPy, and Node 22+ for JavaScript replay. No compiler, GPU or LLM weights are needed. Fetch the directory's contents from the Space repository; keep `core.mjs` in its parent directory for the independent Node test.

```sh
python3 prepare_reproduction.py
OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 python3 experiment.py
python3 control_tests.py
OPENBLAS_NUM_THREADS=1 python3 fresh_stress.py
node test.mjs
```

The Python runs write to `reproduced/`, never over the retained reports. Inputs are pinned public synthetic artifacts. The JavaScript test reads the retained reference outputs. Minor floating-point/timing differences across runtime versions are expected; no timing guarantee is made. Check SHA-256 pins before accepting artifacts. A self-hosted hash manifest is not a digital signature or an independent source of trust.

## Evidence files
`protocol.json`, `results.json`, `control-results.json`, `node-replay-tests.json`, `fresh-stress-protocol.json`, `fresh-stress-results.json`, `serving-profile-example.json`, `training-policy.json`, `proposal-handoff.json`, `axis-coverage.json`, `flashprove-preflight.json`, and `limitations.json`.

`models/refitted_head.json` is the actual fitted candidate checkpoint. Quantization variants and shuffled-target negative-control weights are also retained. `.bin` files contain packed numeric payloads decoded by `experiment.py`; they are not standalone safetensors models or packed integer kernels.

## Integration boundaries
The Academy-compatible pack awards local practice points only; it is not a canonical Academy import. The 34-axis register is relevance coverage, not 34 measurements. The existing 100-candidate inventory remains metadata/README screening, not a completed source audit. The proposal packet is suitable for review and SovSpace projection but does not mutate the shared graph. EAT content and quiz answers do not enter training automatically.

FlashProve preflight did not pass the available Linux compiler requirement; no cryptographic proof was generated or verified. We did not reproduce StrataFlux or transformer KV-cache divergence. All new evidence is unsigned research and not admitted to the canonical GSPC board.

## Licensing
New research wrapper and browser code use the parent Space's MIT license. The reused Memory Venturi source retains the exact upstream MIT license in `vendor/LICENSE` and its pinned canonical revision in `control-results.json`. The original synthetic-model/data notices remain in the parent model/data directories. External papers are linked and discussed, not redistributed as licensed training material.