Council of AI / MEOK · 11 November 2026
Verified work.
One coherent workspace.
What runs, what was tested, what failed and what comes next. Explicit calculation is a useful capability—not evidence of a smarter foundation model.
Open private MEOK · sign-in requiredRead the methods and release planQualification: Qualified for the stated arithmetic and workflow scope. Counts, source identities and limits are recorded in release.json. Finite checks cannot guarantee all future behaviour. Unit/UI providers are substituted; live CPU/MCP calls are separately recorded.
Runs on your device · no model or calculation API request
Try exact arithmetic
Enter decimal numbers, +, −, ×, ÷ and parentheses as + - * / ( ). This browser subset returns reduced fractions. Python/MCP additionally supports bounded assignments and listed functions. Neither interprets arbitrary prose automatically.
Ready. No model call is required.
Inspect the local calculation record
No calculation yet.
Editing clears the old result. Delayed digest results are bound to their original input. Records are unsigned and grant no authority.
Master consolidation
Keep useful components. Keep their boundaries.
Shared task state and evidence can connect different workers without one unrestricted process. Existing components are linked rather than cloned. Remote protocols belong at useful service boundaries, not around every matrix operation.
Conversation runtime
Pinned Cosmos chat. No new model-quality gain is claimed.
Source and evidence →Exact arithmetic
/calc evaluates the expression, not the correctness of a word-problem interpretation.
Source and evidence →Own process predictor
980-parameter synthetic state predictor; not a general foundation model.
Source and evidence →Six-checkpoint replay
Analytical head fit and precision variants; not promoted to the chat worker.
Source and evidence →Specialist receiver
Compatibility and display binding are not truth or human authorisation.
Source and evidence →GSPC axes
Only three classification banks were run in the earlier pilot; not 34 new measurements.
Source and evidence →Memory Venturi
Revision-bound cache reuse is not permanent memory or weight unlearning.
Source and evidence →Optional saved copies
Source is qualified; account isolation, actual storage and restart/reopen need hosted acceptance.
Source and evidence →Academy
Practice paths do not establish real learner gains or appointed reviewers.
Source and evidence →Photonic bridge
No physical optical processor was used in this release.
Source and evidence →SWE-bench
No repository-repair score. HumanEval is a different task.
Source and evidence →The model-to-program candidate did not earn promotion
The frozen 10 November comparison asked the unchanged Cosmos chat for direct answers and arithmetic programs. None of the 33 returned program proposals met the declared language. A calculator cannot legitimately repair an invalid proposal by guessing what its author meant.
| Split | Completed pairs | Direct correct | Program + calculator correct |
|---|---|---|---|
| Development | 16 / 16 | 4 / 16 | 0 / 16 |
| Fresh test, partial | 17 / 48 | 4 / 17 | 0 / 17 |
67 calls were attempted: 66 returned and one hit the free GPU quota. The run paused for an unrelated source change and resumed only unattempted cases after six execution files matched. No paid fallback, answer repair or retry was used. This incomplete experiment is not a full benchmark or an independent Council ranking.
Released: explicit /calc 0.1 + 0.2 in private chat. Not released: automatic prose-to-calculation routing. The hardened calculator does not rewrite the retained experiment's scores.
Work backwards from useful outcomes
Dependency-ordered next steps
| Workstream | Next implementation | Required evidence |
|---|---|---|
| Availability | Keep CPU tools separate from provider quota and report failures explicitly. | Source/serving identity matches; no silent paid fallback. |
| Task Contract | Compare direct and wrapped instructions on separate development data. | Frozen output contracts and no reference-answer leakage. |
| Capability | Compare appropriate general/coding workers under matched budgets. | Correct/all-items, format coverage, refusals, latency and actual cost. |
| Persistent Memory | Qualify opt-in save/list/restore/remove, account isolation and restart/reopen. | Actual mounted-storage readback; temporary storage is not called durable. |
| Swe | Qualify an isolated repository editing/testing agent and official harness. | Original fail-to-pass and pass-to-pass results; every attempt counted. |
| Fine Tuning | Train only on eligible, rights-cleared development episodes. | Fresh evaluation and prior-skill retention; no promotion from loss alone. |
| Physical Backend | Map a profiled bottleneck onto provisioned hardware. | Transfers, calibration, accuracy, latency and energy measured end to end. |
| Canary Release | Release one reviewable version and verify published journeys. | Rollback artifacts; consent, refusal and integrity checks; gaps stay visible. |
Clearer contracts → better proposals → verifiable outcomes → eligible experience → qualified learning. Each arrow is a hypothesis or engineering dependency, not automatic compounding.
Inspect and reproduce
Download tested source and evidence Post-publication qualification · Numerical receipt verifier
Public files contain authored code, synthetic examples, aggregate research and references—not private conversations, credentials, storage objects or answer keys. Hashes bind bytes; they do not prove factual truth or human authorisation. Benchmark test sets remain excluded from training.
Artifact hashes · Component and dependency map · Python source · Python tests · Browser source · Browser tests · Reproducer · Full methods