README.md

metadata

title: CMF Decision
emoji: ⚡
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
short_description: Try resonance-based decisions from one portable CMF file
license: apache-2.0
models:
  - infosave/cmf-decision
tags:
  - cmf
  - cortiq
  - jev
  - decision-making
  - semantic-routing
  - intent-classification
suggested_hardware: cpu-basic
pinned: false

CMF Decision — live demo

Enter a request, choose a skill and inspect the decision, confidence gate and reconstruction errors. The model selects a label or abstains; it does not generate a text answer.

Download the model · Use the hosted routing service · Train your own skill

What runs here

• The published Cortiq 0.8.0 Rust binary and the unmodified 304.5 MB CMF.
• Three bundled skills: BANKING77, CLINC150 and MASSIVE, with English examples.
• Free CPU hardware. Local installations also support Metal and Vulkan.
• No oracle, API keys, online learning or calls to paid inference providers.
• Reconstruction scores and timings come from the actual engine, not fixtures.

Requests run on the Hugging Face host, not in your browser. Do not submit sensitive information. The demo does not log request text or use it for training. It exposes only the demo endpoint, not the engine's administrative API.

The displayed model time is encoder + reconstruction time, excluding the browser, network and web-server overhead. Shared CPU demo timings are not the GPU benchmark.

Run locally

docker build -t cmf-decision-demo .
docker run --rm -p 7860:7860 cmf-decision-demo

The build downloads a checksummed release binary and a revision-pinned, checksummed CMF file. No Rust build, PyTorch or extra Python packages are needed.

For a local native binary, set CORTIQ_BIN and CMF_MODEL and run python3 app.py. CMF_DEMO_HOST defaults to loopback locally and 0.0.0.0 in the Docker image. PORT defaults to 7860. Tests: python3 -m unittest discover -s . -p 'test_*.py'.

The underlying component notices are included at /home/user/app/NOTICES.md in the Docker image. Model/source documentation lists benchmark protocols and licenses.