# Serving Configuration as Evidence
## Precision, cache identity and bounded head adaptation in MEOK

**CSOAI engineering research note · 10 November 2026 · v0.3**

**Status:** AI-assisted, reproducible engineering research. Not peer-reviewed, not a customer case study, not a production release, and not admitted to the canonical GSPC board.

## Abstract

We extend the existing MEOK synthetic process-model experiment without modifying its production defaults or the separate Cosmos LoRA experiment. Six checkpoint variants are compared on 560 previously published test transitions and 640 newly generated transitions after candidate freezing. Symmetric INT8 and INT4 weight representations alter the highest-scoring predicted state on 10 and 114 of the original 560 cases, respectively. A 68-parameter output head, fitted on the original training split while freezing 912 backbone parameters, reduces mean absolute error from 0.28797 to 0.27495 percentage points on the original test bank and from 0.18738 to 0.16853 on the fresh bank. A large original-bank worst-case error remains. A configuration-bound bridge to the existing Memory Venturi cache passes synthetic isolation controls; deliberately supplying a stale worker identity reproduces stale reuse. Thirty-six control checks and 38 independent JavaScript replay checks pass. These results support narrower, configuration-specific evidence records, not a claim of general intelligence, complete system safety or automatic model promotion.

## 1. Research question and relationship to existing work

The question is not whether all models should be merged. It is whether a small, existing learned component can be improved and evaluated while keeping request identity, numerical implementation, evidence, memory, human authorization and deployment decisions separate.

The starting point is `csoai/meok-process-world-v0`, a 980-parameter synthetic procurement-state predictor. Its canonical source is Council PR #4176. Its earlier constrained variant improved average error relative to the raw network, but the published worst-case error remained substantial. Its deterministic generator has known mechanics and fully observed profile counts: the exact rule remains the best comparator for that narrow task.

This note extends, rather than replaces, the earlier Memory Venturi white papers and the existing 100-candidate discovery inventory. The inventory is metadata/README screening, not 100 code audits or 100 installed integrations. The new public interface reuses the prior inference implementation and the existing Memory Venturi code. The private MEOK runtime, shared M4 worktree, canonical board, signing process and approval owners are not overwritten.

## 2. Fixed assets and experimental design

Model repository revision: `90d6aa6c87629b9f4296142ba2019247b51686c4`.

Dataset repository revision: `4a522537c9eb1522a3aad1432e9626e72e3a2733`.

Original weights SHA-256: `9fbffae6eb2f025a729a74022d041437f630bf06d60b4964019be10b70ba7c8b`.

The original episode-disjoint split has 1,600 training rows, 400 validation rows and 560 test rows. Inputs contain four previous-state proportions, three profile proportions, a revocation flag, expiry information, an evidence-awareness flag and time. Outputs are four next-state proportions. No customer records, human review answers, proprietary prompts or real-world demographic data enter the experiment.

`protocol.json` records the candidate family and selection rule. The backbone remains frozen. Five ridge penalties are compared using validation error only: 0.000001, 0.001, 0.1, 1 and 10. The 68 output-layer coefficients and biases are fitted analytically; this is actual parameter fitting but **zero gradient optimizer steps**. The selected penalty is 0.000001. A shuffled-target control uses seed 17 and the same fitted family.

The original test bank is already public, so its replay is not called a new blinded test. After all six candidate files were frozen and hashed, episodes 320–399 were generated using the pinned original engine, producing a second bank of 640 transitions. No candidate is selected or refitted against that bank. It is a fresh same-generator stress test, not independent real-world validation.

The fit and Python checks ran on the existing authorized M2 CPU, not a new Hugging Face training job. The Hugging Face page serves the artifacts and runs the independent replay on the visitor's browser CPU. The independent JavaScript calculation matched the Python metrics within 2.22e-16.

## 3. Precision results

The original numerical weights are compared with float32 matrices and symmetric per-layer INT8/INT4 matrices. Biases and scaler parameters remain float64. Packed matrices are decoded into ordinary float64 arithmetic; **no packed integer kernel, transformer KV cache, GPU streaming or StrataFlux implementation is exercised**.

All results below apply the original clipping, normalization and expiry constraints. Errors are population-fraction errors expressed as percentage points.

| Variant | Original 560: mean / worst error | Fresh 640: mean / worst error | Top-state changes vs original: old / fresh |
|---|---:|---:|---:|
| Original float64 | 0.28797 / 74.99077 | 0.18738 / 6.17784 | 0 / 0 |
| Float32 weights | 0.28797 / 74.99077 | 0.18738 / 6.17783 | 0 / 0 |
| INT8 weights | 0.32907 / 72.92649 | 0.24613 / 6.98325 | 10 / 13 |
| INT4 weights | 4.50944 / 26.61762 | 4.89917 / 38.62462 | 114 / 145 |
| Fitted output head | 0.27495 / 74.85274 | 0.16853 / 5.26079 | 7 / 7 |
| Shuffled-target control | 5.15274 / 42.58192 | 5.51680 / 62.86073 | 183 / 198 |

A top-state change is a change in the largest of four predicted proportions, with deterministic first-index tie handling. It is not a measured chat-answer change, executed tool action or human outcome. Float32 changes small numerical values but produces no top-state changes in these two banks.

INT4 illustrates why a single score is inadequate: it reduces the worst error on the old bank while making average error much worse, and it is also worse on the fresh bank. A smaller representation is not automatically a better model.

The packed numeric payloads are 8,016, 4,304, 1,544 and 1,080 bytes for float64, float32, INT8 and INT4. These figures exclude manifests, JSON representations, datasets and the application. Five local timing repetitions per representation retain warm and decode-plus-inference measurements. They are diagnostic only: cross-variant order is not randomized, the matrices are tiny, and neither GPU transfers nor energy were measured. No production speed or energy advantage is claimed.

## 4. What the fitting experiment establishes

The fitted head improves average error by approximately 4.52% on the original bank and 10.06% on the fresh bank, relative to the constrained original. The old-bank worst error remains approximately 74.85 percentage points. It would be misleading to promote the model using the average alone.

Paired episode bootstrapping uses 10,000 resamples, preserving all eight transitions within each episode. The exploratory 95% interval for head-minus-original mean error is [-0.02070, -0.00570] percentage points on the original bank and [-0.02965, -0.01006] on the fresh bank. These intervals characterize the sampled synthetic episodes, not uncertainty about deployment populations or all possible model-selection procedures.

The shuffled-target control performs substantially worse, and the exact deterministic generator still wins on its own known task. The artifact is therefore retained as a **research candidate**, not a new default. This is bounded supervised adaptation of a small component, not recursive self-improvement or ASI.

## 5. Serving identity and cache correctness

The bridge reuses `harness/meok-memory-venturi/memory_venturi.py` at canonical source revision `4f26e22ef66e232a8af8fffb89e44e898cd747ea`. The public source bytes match the inspected private runtime copy. The new serving profile binds:

- concrete model-artifact hash and precision;
- runtime and method hashes;
- source snapshot, cache mode and generation;
- policy epoch, with tenant/principal/current authorization still checked by the existing host contract.

A negative control intentionally changes the true numerical variant while retaining the old worker identity. The existing cache then reuses the wrong prior result. Supplying the accurate variant-specific identity forces a fresh result. Source corrections, host policy changes, principal changes, truncated dependency impact and authorization withdrawal are also tested.

This is not a cache-library exploit or a claim that the library can discover an undisclosed substitution. **The authenticated host must supply truthful serving identity.** A digest binds reported information; it does not independently establish that the report is true. An attestation or compatible inference proof can strengthen a different part of that trust chain.

This answer-cache experiment is distinct from the published transformer-prefix-cache study [1]. The latter investigates numerical differences between cached and recomputed KV paths. Our feedforward MLP has no KV cache. The common architectural lesson is to record material serving state rather than treating visible request bytes as a complete execution identity.

## 6. FlashProve and StrataFlux: opportunities, not reproduced claims

Anthropic released FlashProve on 2 November 2026 [2,3]. Its purpose is to verify that a response arose from a committed model computation. That does not establish that the response is factually correct, that its training data were licensed, or that a human authorized its effects. A third-party API response cannot acquire a valid inference proof merely by being wrapped in our receipt. Compatible proof generation and trusted public-input binding remain necessary.

The existing Linux host has g++ 11 and limited available memory; the documented CPU-verifier route requires g++ 13 or later. No verifier was built, no stored proof was verified, and no MEOK proof was generated. The preflight failure is retained. Upstream labels the release a research artifact that is not maintained or accepting contributions; reuse must preserve its Apache-2.0 and other applicable notices.

StrataFlux was submitted on 5 November 2026 [4]. Its selective-precision approach co-designs additional precision delivery with computation and reuse. It motivates measuring quality, bytes, delivery, decoding and full execution together. Our decoded-MLP quantization is not a reproduction of its GPU/RAM/SSD pipeline, reported throughput or language-model quality results.

## 7. Human review, Academy, SAFE and the 34-axis index

The release includes eight evidence-linked practice questions in the existing question/options/answer/explanation/xp_reward shape. The browser can grade them for up to 160 local practice points. No persistent Academy XP, authenticated identity, certification, approval power or consent to train on responses is granted. The pack is not automatically imported into the canonical Academy.

Training eligibility checks hold evaluation splits, unknown rights, private/unknown data and unverified provenance. EAT ingestion is not training permission. Negative results may become review lessons; they do not become unconsented training examples by default.

The retained canonical roster contains 34 axes. `axis-coverage.json` records research relevance or not-tested status for every axis. It does not turn 34 labels into 34 new measurements. A fresh public board read returned HTTP 403; the exact retained snapshot and that limitation are named. Existing SAFE exchange material is an internal reference proposal, not an official schema or submission. The inspected RAS bridge remains proposal-only; the expansion of the user's SAS-to-RAS term was not established, so no authority transition is invented.

A useful three-role arrangement is proposer, challenger and verifier, with human authorization separate. It is not automatically Byzantine fault tolerance. Under the assumptions of a correct authenticated BFT protocol, 4, 13, 33 and 34 independent replicas can use conservative quorums of 3, 9, 23 and 23 for fault budgets 1, 4, 10 and 11 [5]. Only this arithmetic is tested here. A central coordinator counts only if it is genuinely an independent voting replica. Correlated clones and model votes do not establish truth.

## 8. Legacy boundaries and downstream reactions

The added legacy controls preserve 31-digit decimal strings and negative zero, refuse binary floats or inconsistent scales, and check a CP037 byte roundtrip. They complement existing legacy-codec work; they are not a COBOL compiler, copybook implementation, bank interface or settlement test.

A proposal packet connects source/model changes to precision evidence, serving profiles, cache checks, Academy lessons and human review. This is a dependency projection for SovSpace, not an installed replacement for its shared graph. Dream proposes; Evolve compares; reviewers decide whether a candidate is admissible. None of those steps silently grants execution, payment, signing or GSPC admission.

The practical commercial candidate is a reproducible serving-fidelity assessment: specified workload, exact configurations, error and resource accounting, counterexamples and a reviewable report. This release proves the existence and bounded behavior of the research artifacts. It does not prove customer demand, revenue, enterprise readiness or superiority across industries.

## 9. Reproduction and retained limits

Use `README.md`, `prepare_reproduction.py`, the three Python scripts, and `test.mjs`. All source pins, input hashes, selected and failed candidates, exact results and limitations are retained. The public browser replays 7,200 inference evaluations across the two banks. Thirty-six control checks and 38 Node replay checks pass; these are not whole-platform certification.

No new paid HF job or hardware was started. The separate Cosmos adapter and its majority-label failure remain unpromoted. The private runtime and shared worktrees were not changed. FlashProve execution, real transformer cache experiments, a distributed BFT deployment, authenticated approval-to-effect, customer validation, complete SovSpace installation, all 100 code audits and all 34 formal axis measurements remain outside this completed bounded experiment.

## References

[1] Aditi Patodiya. *Same Request, Different Answer: Quantization Amplifies Cache-Induced Divergence in LLM Serving*. Submitted 4 September 2026. https://arxiv.org/abs/2609.04748 . Mechanism and reported results were researched, not reproduced here.

[2] Anthropic. *Progress on provable inference for large language models*. 2 November 2026. https://www.anthropic.com/research/provable-inference .

[3] Anthropic. *FlashProve*, research artifact v0.16.0, README and verifier requirements. https://github.com/anthropics/FlashProve .

[4] Jake Hyun, Ziyu Zhao, Jae-sun Seo and Mohamed S. Abdelfattah. *StrataFlux: Memory-Hierarchy-Native Low-Bit LLM Inference*. arXiv:2611.08363v1, 5 November 2026. https://arxiv.org/abs/2611.08363 .

[5] Miguel Castro and Barbara Liskov. *Practical Byzantine Fault Tolerance and Proactive Recovery*. ACM TOCS 20(4), 2002. https://doi.org/10.1145/571637.571640 .

[6] CSOAI. Original model, synthetic dataset and canonical research source. https://huggingface.co/csoai/meok-process-world-v0 ; https://huggingface.co/datasets/csoai/meok-process-world-transitions-v0 ; https://github.com/CSOAI-ORG/councilof-ai/pull/4176 .

[7] Hugging Face. Storage Buckets and Spaces storage documentation. https://huggingface.co/docs/hub/storage-buckets ; https://huggingface.co/docs/hub/spaces-storage . Buckets are mutable storage, Spaces host applications; neither is inherently an offline platform or an immutable evidence archive.