CKODMKby MFENX

CKODMK 0.5.1

Verify a model before it runs.

CKODMK checks a signed model package, binds it to its source and contract, replays the required behavioral checks locally, and blocks model admission when any required condition fails.

Execution
In this browser
Files
Remain on this device
Runtime
ONNX Runtime Web
Offline
Available after installation

Admission gate

Proof-carrying model

The candidate cannot be downloaded from this gate until its issuer signature, RootID, file bindings, validity period and required behavioral checks pass.

Admission resultSelect a PCM or run the included trained-model package.
Ready
  1. 1
    AuthorizationIssuer, lineage, lifecycle, claims, RootID and artifact bindings
    Waiting
  2. 2
    Runtime replayEvery required behavioral contract check
    Waiting
  3. 3
    AdmissionUnlock only the candidate that passed
    Locked
Statement RootID
Not verified
Issuer
Not verified
Valid until
Not verified
Candidate digest
Not verified
Model lineage
Not verified
Release authorization
Not verified
Model scale
Not verified
Native parameters
Not verified
Precision
Not verified

Workspace

Model comparison

Nothing is uploaded. Closing the page clears the active session unless you download the files.

1
Start with a working modelRun an included trained model, or continue below with your own files.
2

Files and settings

Supply a source model and dataset. Build a candidate here or select one you already have.

4

Measure this browser

After the comparison passes, CKODMK alternates source and candidate runs and records the observed timing. It does not record the device model, operating system, browser version, account, or network address.

Complete the comparison first.Locked
Source p50
Not run
Source p95
Not run
Candidate p50
Not run
Candidate p95
Not run

Reference runs

Results from included models

These records make the examples inspectable before you run them locally. Select a row to view its measurements.

ModelNew Optdigits CNN / Static INT8
Recorded
Accuracy
96.994992% to 96.994992%
Decision changes
0 / 1,797
Bytes
8,749 to 3,865
Size ratio
2.263648× smaller
Sample p95
0.040512 to 0.053696 ms
Maximum finite set L∞
0.336463928

Trained in this repository on the official Optdigits training split. The candidate is smaller and made the same classifications on the complete official test split. Its recorded latency is higher.

Command line

Use CKODMK in a build pipeline

The browser provides local PCM admission. The Python package creates issuer keys, signs PCM files and verifies cryptographic bindings for build systems. The source repository is private.

Install version 0.5.1python -m pip install 'mfenx-ckodmk[production] @ https://mfenx.com/ckodmk/downloads/mfenx_ckodmk-0.5.1-py3-none-any.whl'
Package checksum and supported file profile

Verify the package

curl -O https://mfenx.com/ckodmk/downloads/mfenx_ckodmk-0.5.1-py3-none-any.whl.sha256

Supported browser files

  • Source and candidate ONNX files up to 64 MB
  • One Float32 classification input and one output
  • NPZ with Float32 inputs and Int64 labels
  • Batch size one, up to 25,000 rows and 250,000 output values

Large checkpoint trees

  • Safetensors and weights-only PyTorch state dictionaries
  • Single-file or sharded checkpoints with exact index coverage
  • Streaming file and tensor hashes without whole-tree buffering
  • Use inspect-checkpoint-scale from the command line