# MEOK: consolidated components and qualified improvement

**11 November 2026 · Author-run product engineering and research**

## Objective and release boundary

The objective is a useful end-to-end workspace: receive a task, maintain the distinction between evidence and prediction, choose an appropriate operation, return a checkable result, and preserve enough provenance to correct it. Success is correctly completed, authorised work under an explicit cost and latency budget, not the number of models, messages or tools connected.

This release adds exact arithmetic to the existing private MEOK conversation and MCP service, and publishes its authored implementation and a smaller browser demonstration in the existing public research lab. It preserves the current model, conversation, optional saved-copy source, Academy, Memory Venturi, specialist receiver and previous research experiments. No foundation-model weights changed. No physical photonic processor, new paid job, protected production merge, human appointment or neutral public GSPC self-ranking is claimed.

The public demonstration is not the private conversational model. It performs the displayed calculation locally, without credentials, a model or a calculation API request. The private service remains private.

## Two separate arithmetic problems

Calculating an expression and choosing the correct expression for a natural-language problem are different tasks. The new component addresses the first. It does not establish the second.

The retained 10 November candidate trial selected 16 development problems and 48 fresh GSM8K test problems, excluding the previous capability pilot's test rows. It froze two recipes before generation: direct numeric answers through the existing chat, and arithmetic-program proposals through the same chat followed by a bounded calculator. Model settings were unchanged, and answer keys were stored separately from prompts.

Of 128 planned calls, 67 were attempted and 66 returned. The first phase stopped when another workstream changed the serving revision. Six execution files were byte-compared before continuing only unattempted tasks. The second phase ended at an actual free ZeroGPU quota refusal. No failed generation was silently retried through a paid provider or another identity.

All 16 development pairs and 17 of the planned 48 test pairs completed. Direct answers were correct on 4/16 development cases and 4/17 completed test cases. None of the 33 returned program proposals satisfied the declared calculator language. The program-plus-calculator recipe therefore scored 0/16 and 0/17, respectively. Invalid proposals were not omitted from the denominator or converted to correct labels after inspection.

The complete confirmation set did not finish, and automatic word-problem routing was not promoted. This is evidence of failure of that particular proposal/interface recipe, not a proof that program-assisted models are generally ineffective. Extracting a convenient number from the prose would be a different evaluation method. The original scores remain frozen and are not recomputed with this release's hardened calculator.

The protocol and response hashes are retained in ledger.json. Public content includes aggregate own-product research, not private conversations or benchmark answer keys. This is not an independent Council ranking of its own system.

## Exact-calculation implementation

The Python component accepts one expression or a bounded sequence of single assignments ending in `answer = ...`. Numeric literals are converted from decimal source text into exact rational values. The supported arithmetic operations and functions are explicitly enumerated in the source. Returned fractions are not silently rounded into decimal approximations.

There is no Python `eval` or `exec`. Imports, arbitrary calls, attributes, file access, loops, comprehensions and function definitions are outside the accepted language. The implementation bounds UTF-8 input size, AST nodes, expression depth, statement count, literal length, rational numerator/denominator size and exponents. `round` uses half-to-even semantics. Refusals return no calculated result.

Review identified a resource risk in scientific notation: a huge negative exponent can underflow to finite floating-point zero during parsing, while a subsequent rational conversion can attempt an enormous denominator. Version 1.1.0 checks the textual exponent before constructing the Fraction. The regression test substitutes an allocation sentinel; it does not construct the dangerous denominator. The original experimental calculator remains untouched so its study provenance is preserved.

The private command is explicit: `/calc 0.1 + 0.2`. Existing consent and the tool switch still apply. Natural-language maths is not automatically rerouted. Images attached to a text command are refused rather than silently ignored. The tool returns the actual expression, exact result or refusal, and a record made with the existing GSPC event implementation. The event is unsigned and non-authorising; it records a function return, not independent certification of the assumptions in the expression.

The public browser component has a smaller contract: decimal literals, +, -, *, / and parentheses. It uses reduced BigInt numerator/denominator pairs and does not implement the Python assignment/function language. Editing the input clears its previous result. A generation counter and exact-input comparison prevent a delayed hash operation from attaching an old result to changed input.

## Qualification and failure accounting

The publication sequence is source pinning, parallel-work reconciliation, isolated additive candidate, focused tests, composed regression, source-manifest verification, exact-parent publication, serving-revision readback, real API/MCP calls and rendered user journeys.

Twenty calculator test methods include exact decimal arithmetic, bounded resource cases, rejected constructs and 160 seeded arithmetic parity cases. Thirteen additional integration tests exercise the real conversation adapter and existing event implementation. The focused combined suite passed 86 methods. The composed runtime passed 309 methods, including retained memory, Academy, specialist and conversation checks. Expensive model providers are substituted in those unit/UI tests; the count is not a model accuracy result.

An initial composed run failed the retained OAuth login-button test because the local cached environment lacked declared OAuth dependencies. That failure is retained. The dependency environment was corrected, rather than weakening the test or modifying the unrelated memory service. A subsequent stalled test process was stopped; a bounded rerun with Gradio analytics disabled completed. The initial failure and interruption are not represented as clean first-attempt success.

Actual hosted readback exercised five chat requests and six MCP requests. All thirteen grouped API assertions passed: access and consent, exact arithmetic, refusal/no result, help and retained axis lookup, six vendor-independent arithmetic outcomes, existing event validity, unsigned status, no model calls and exact serving revision. The browser arithmetic tests contain 188 assertions. Rendered browser qualification is recorded separately after execution; CodeMirror virtualisation and the known private-host HF badge warning must not be relabelled as model defects or silently erased.

Finite tests do not guarantee every future input, every supported browser or every failure mode. No universal 100% correctness, independent security certification or hardware performance claim follows from these checks.

## Consolidated components, without duplicate machinery

The hosted conversational worker and the published 980-parameter process predictor are different models. The predictor is trained for a narrow synthetic transition task; its worst errors remain visible. The existing six-checkpoint browser replay covers an analytical output-head fit and precision variants, not a new general model or chat upgrade.

The specialist receiver checks submitted artifacts against an explicit compatibility contract. Another workstream repaired the presentation so earlier success is not shown as the result of a changed or refused request. This release links that published repair instead of duplicating it. Compatibility is not truth, independent review or permission to execute.

Memory Venturi explores revision- and dependency-bound reuse. A compact root can navigate larger evidence, but it does not eliminate supporting storage. A content hash binds bytes, not honest provider identity; misleading identity metadata remains misleading when hashed. Invalidating external evidence or a cache is not proof of removing information from trained weights.

The optional saved-copy implementation requires explicit opt-in, current account identity and configured persistent storage. Hosted save/list/restore/remove, account isolation, restart/reopen and actual storage readback remain separate acceptance gates. An earlier failed mount association was removed without deleting the bucket or its objects. That recovery was an availability repair, not proof of durable memory. No temporary fallback should be described as permanent storage.

Academy paths and practice cases support structured review. They do not establish real learner gains, an appointed professional council or coverage of every industry. Likewise, the GSPC 34-axis interface is not thirty-four newly trained specialists or thirty-four new measurements in this release.

The Q.ANT bridge exercises the official CPU software interface. Classical photonic computation, co-packaged optical networking and quantum computation remain different tracks. A CPU or circuit simulation must not be relabelled as measured physical acceleration.

## Work backwards: next moves and their dependencies

First protect availability and explicit failure reporting. CPU operations should not require GPU quota, while a model-provider failure must not trigger an undisclosed paid fallback. Retain a rollback source and separate CPU, model and storage health.

Then isolate the task contract. Compare the same model with direct and wrapped prompting on development cases, with an appropriate output budget and output-format validation. A general assistant prompt can conflict with a strict task response. This is a hypothesis to test, not a justification for rewriting the old benchmark scores.

Next compare suitable general and coding workers under the same prompts, tools, generation budget and grading. Report correct/all-items, format coverage, abstention, latency and actual cost. A published score under another protocol is context, not a controlled head-to-head result. Preserve regressions and tail errors rather than optimising only the mean.

Repository repair requires an isolated editing/testing executor and official tests before a SWE-bench score exists. HumanEval function synthesis is not repository repair. Pin the environment and allowed edit/test budget, keep reference patches inaccessible to the agent, and record every attempt.

Fine-tuning is later in this chain. Admit only rights-cleared development episodes with source and outcome provenance, task quality checks and appropriate consent. Do not turn inspected public benchmark test answers into training data. Separate episodic facts, reusable procedures and stable trainable capabilities. A new adapter must beat its unchanged parent on fresh evaluation, retain previous capabilities and survive review; lower training loss alone is not admission.

For photonic work, first identify a useful profiled bottleneck and supported mathematical operation. After legitimate hardware access, include preparation, conversion, transfer, calibration, compute and readout costs. Faster arithmetic does not automatically improve an end-to-end workflow dominated by other stages.

The proposed chain is clearer contracts → better proposals → verifiable outcomes → eligible experience → qualified learning. Each step is a measured hypothesis or engineering dependency, not automatic compounding. The machine-readable plan remains acyclic and records acceptance conditions for each step.

## Reproduce and inspect

Run `python reproduce.py` from the downloaded directory to check the manifest and new Python tests. Run `node test_calculator.mjs` for the browser arithmetic subset. Serve index.html over HTTP(S) for the visual demonstration. The private integration is separately versioned and requires account authorisation; no credentials belong in the public demo.

Primary technical references: Python Fraction semantics (https://docs.python.org/3/library/fractions.html), AST structure and resource limits (https://docs.python.org/3/library/ast.html), Hugging Face MCP hosting (https://huggingface.co/docs/hub/en/spaces-mcp-servers), and mutable bucket storage (https://huggingface.co/docs/hub/storage-buckets). These documents support the design mechanisms, not an endorsement or independent validation of MEOK.

## Completed on M4: final publication gate

The current release passed 309 composed test methods, including actual serving-queue validation. Expensive providers are substituted in those tests. Separately, the exact deployed runtime passed 13 API/MCP assertions across five real chat-tool requests and six MCP requests, eight private-browser assertions and nineteen anonymous public-browser assertions. These cover calculation, refusal, consent, current-input result binding, reset, mobile width and published navigation. No model inference or stored user-conversation write was used in these CPU acceptance checks.

A post-deployment startup failure exposed a missing qualification step: an unqueued JavaScript-only event cannot be cancelled through the server queue. That configuration was removed, the source check was added, and the actual service was reverified after repair. Input changes now clear the old arithmetic result through the existing change-event pattern; returned results remain bound to the expression. The original failure is retained rather than counted as a successful deployment.

The private-host badge warning remains documented. The public browser run reported no JavaScript errors. The earlier M4 disk-full interruption is retained; this pass first verified free space before qualification.

A separate synthetic model-chat availability request encountered the free ZeroGPU quota; no alternative identity, undisclosed provider, subscription or paid runner was used. This does not affect the demonstrated CPU arithmetic path, but a RUNNING Space alone must not be advertised as unlimited model availability. It is not a new accuracy measurement.

The newer public numerical receipt verifier was integrated by link, not duplicated. Its 37 controls and fresh-publication reproduction are recorded by its existing workstream at revision e381bcac01b0de1379fce326226a27d5f8e8462f. Its recomputation is stronger than checking a checksum alone, but does not authenticate source capture, a human or an external provider. Its tests are not added to this release's 309 count.

The downloadable source bundle excludes credentials, private conversations, raw private runtime records and benchmark answer keys. It includes the authored calculator, browser implementation, source manifest, methods, aggregate experiment, qualification summary and dependency-ordered work plan. The existing public model, datasets, receiver and numerical-verifier files are preserved.