11 November release

Exact arithmetic, component map and release qualification · Methods and next steps

Authored research and explicit calculation; not a new general-model performance claim.

COUNCIL OF AI / MEOK RESEARCH

Specialists together.
Evidence kept separate.

RESEARCH v0.2

Real learned inference, explicit rules and a reproducible comparison. One small model—not a merged foundation model.

980 trained parameters560 test transitionsBrowser CPU onlyNo new training
11 November · verified research release

Open the consolidated journeys and evidence. Current-result display is cleared when inputs change or fail; the existing models and test banks are preserved.

New v0.3 research extension

Run the precision, fitted-head and evidence-learning batch · Read the new research paper. The original v0.2 comparison remains below.

What this proves—and what it does not

This checkpoint predicts a synthetic procurement process. Exact rules know the generator and outperform the learned model on this task. Neither result establishes real-world safety, general intelligence or permission to act. Cosmos is not loaded.

Checking pinned weights and dataset…
01 / RUN A TRANSITION

Compare the components

No model-provider request is made by this lab

Previous A/B/C/undecided; profile 0/1/2; A revoked; ticks until expiry ÷ 8; evidence-aware flag; tick ÷ 8.

What each component returns

ComponentABCUndecided
Load a row and run the model.

The learned network is advisory. The exact rule is selected only inside this known synthetic process.

Inspect the full comparison record
No result yet.
02 / REPRODUCE THE EVIDENCE

Replay all 560 held-out rows

The test bank spans 70 disjoint episodes from the original experiment. This is a replay of a published bank—not a fresh blinded test. Errors below are population-fraction errors, expressed in percentage points.

ComponentMean absolute errorWorst absolute errorOutputs outside [0,1]
Run replay to calculate the values on your device.

A lower average error does not remove large failures. The constrained learned model still has about 75 percentage points of worst-case error in this bank.
Inspect the replay record
Not run in this tab.
03 / REUSE, DO NOT DUPLICATE

Existing MCP runtime and pinned sources

The private MEOK workspace already serves the same process predictor through MCP. This public interface computes locally; it does not send a credential or silently call the private server.

MCP integration

meok_hf_runtime_predict_process runs the pinned process model. meok_hf_runtime_compact_root produces a content-addressed index, not durable storage.

Open existing private MEOK workspace ↗

Read connection and authority boundaries ↗

See retained live-check evidence.

Check the source

Original model and limitations ↗
Original synthetic dataset ↗
Canonical model research PR #4176 ↗

Regression tests · Independent engine replay · Artifact manifest

Model and dataset SHA-256 pins