MFENX

MFENX Gate / Local evaluation

Set the conditions.
Test the candidate.

A model release is a decision about particular bytes, a particular source and a particular evaluation contract. Make that decision on your own machine.

Open the local evaluator

CKODMK is the evaluation engine in MFENX Gate. Its browser workspace runs supported ONNX models, compares source and candidate outputs, and replays the contract carried by a signed model package.

Open signed-model admission

No account or upload. The evaluator and its runtime download to your browser; the model and evaluation files you select stay on this device.

What the gate checks

  1. Authorize the package

    The issuer signature, validity period, release authorization and lineage must verify. The model files must match the package’s artifact bindings.

  2. Replay the required behavior

    The engine executes the included evaluation contract locally and checks the candidate’s measured results against its acceptance limits.

  3. Release the exact candidate

    The evaluator unlocks its candidate download only after its required checks pass. Your deployment pipeline must enforce its own corresponding gate before loading a model.

Choose the right evidence

Receipt authentication, artifact inspection and behavioral replay answer different questions. Select the operation that your release policy requires.

Scroll horizontally to read the full comparison.

Operation What you learn Open
Release receipt Whether the issuer’s signed statement is intact, current and trusted. Supply the model separately to check its artifact digest. This operation authenticates a statement; it does not replay the evaluation. Free verifier
Signed ONNX package Whether the package verifies and its required finite behavioral evaluation passes on this device. Replay package
Source–candidate comparison How supported ONNX candidates compare on the evaluation data you select, including measured decisions, accuracy and latency. Local workspace
Large checkpoint inspection The identity and structure of safetensors or weights-only PyTorch checkpoints. Behavioral testing requires a compatible execution adapter and evaluation contract. CKODMK CLI

Read the result in context.

An accepted evaluation concerns the recorded model, data, contract and execution environment. Keep those identities together when a candidate is quantized, converted or packaged for another device.

Start with the published CKODMK reference runs, then run the same workflow with your own acceptance contract. For automation, use the local CLI and integration documentation.