infosave/cmf-decision

CMF Decision

Decisions in milliseconds. One portable CMF file.

Turn a customer message into a clear action: route a request, classify an intent, or choose a tool using your own trained skill. Cortiq selects a label or abstains when uncertain — without generating tokens locally. Connect an optional oracle for unfamiliar cases, then use its answers to grow your local skills.

Local decisions · Custom skills · Optional oracle

Try the live Space → — enter a request and inspect the decision, confidence and reconstruction errors. No installation or API key needed.

Quick start · Jev-compatible requests · Your own skill · Oracle · API · Benchmarks · На русском

Quick start

Use the hosted service

Prefer an API without local setup? allaigate Routing API is a ready-to-use routing service built on the CMF format. Get an API key on the site and follow the integration guide — no model download or server deployment required.

Run locally

Install the latest published CLI from crates.io (requires Rust 1.88 or newer):

cargo install cortiq-cli --locked

Run the same command again to update. Download the model and make your first decision:

curl -fL -o cortiq-decision.cmf \
  https://huggingface.co/infosave/cmf-decision/resolve/main/cortiq-decision.cmf
cortiq decide cortiq-decision.cmf --skill banking77 -p "I still have not received my new card"

Result: card_arrival, answered locally. No cloud key required. Included skills: banking77, clinc150, massive.

Choose your hardware

CPU works without configuration. The same installed CLI also includes GPU support (from Cortiq 0.8.0):

# Apple Silicon
CORTIQ_DECISION_DEVICE=metal cortiq decide cortiq-decision.cmf \
  --skill banking77 -p "I still have not received my new card"

# Vulkan GPU (a hardware Vulkan driver is required)
CORTIQ_DECISION_DEVICE=vulkan cortiq decide cortiq-decision.cmf \
  --skill banking77 -p "I still have not received my new card"

On multi-GPU hosts, also set CORTIQ_DECISION_VULKAN_ADAPTER to a unique part of the GPU name. The same variables work with cortiq serve. GPU guide →

Jev-compatible request format

Cortiq 0.8.5 adds an opt-in adapter for the TypeSafe System One request format. Start it alongside the normal Cortiq APIs:

cortiq serve cortiq-decision.cmf --jev-compatible --port 8080

Point a System One client at http://127.0.0.1:8080/v1/systemone. Its default jev-latest (and the historical typesafe/jev-1.13) is accepted only as a request alias; an omitted model works too. Every response identifies the local CMF model as cmf-decision-0.8.5 — this adapter does not serve or claim to be Jev. GET /v1/models supplies the matching System One discovery shape when the flag is enabled. Request and response details →

From 0.8.6 a question whose options no skill was trained on is learned from the oracle's answers into an auto-skill and then answered locally in milliseconds; labels the oracle rarely picks stay quarantined and keep teaching. How auto-skills learn →

Measured results

Quality: the model alone vs Jev 1.13

With abstention enabled, accepted answers are 97.24–98.70% correct; the model answers 54.30–92.11% of requests locally, depending on the task. Accuracy and coverage together →

Laya: 77 choices on the same Mac

Trained CMF skill: 93.34%, 3.10 ms p50. Laya's base English checkpoint, all 77 label names at once: 35.65%, 917.59 ms p50. All 3,080 rows, both local, no oracle. A 77-option stress test, not Laya's best achievable result: shortlisting and multilingual were not tested, and training conditions differ. Full protocol, failed long-rubric run and raw results →

Speed: now on Metal and Vulkan

1.10–1.24 ms on RTX PRO 4000; 2.28–2.58 ms on Apple M4. The full path is accelerated, not just the reconstruction kernel. All 10,554 test examples keep their CPU decisions and abstentions. No retraining.

M4 uses alternating CPU/GPU requests; RTX uses a warmed continuous stream. Sparse RTX traffic is slower; Metal can have worse tails under desktop load. Protocol, tail latency and memory → · Earlier API comparison with Jev →

Cost: cloud only when needed

Local-only decisions incur $0 in API fees. The optional hybrid run used DeepSeek V4.1 Flash for hard cases, reaching 93.93% / 97.47% / 88.00% overall accuracy. Fees are extrapolated from recorded runs, not a hosting quote. Oracle setup →

Why CMF?

• One deployable file. Encoder, skills and decision rules travel together.
• Resonance, not text generation. Each label tries to reconstruct the input signal; the smallest error wins. A calibrated gate decides whether to answer.
• Your tasks, your control. Add skills from examples; keep existing skill parameters unchanged. Version and roll back learned updates.
• Local by default. Enable an external oracle only when you need one.

Add your own skill

Prepare labeled examples in train.jsonl and describe the labels in question.json:

cortiq decision add-skill cortiq-decision.cmf --skill support \
  --train train.jsonl --question question.json -o support.cmf
cortiq decide support.cmf --skill support -p "Please cancel my subscription"

Example data, calibration and training options →

To serve over HTTP: cortiq serve cortiq-decision.cmf. Requests, keys and router compatibility →

---

Benchmark scope: reused public datasets; CMF trained on train + dev, Jev given two examples per label. The model combines a Cortiq NVG-modified text encoder with trained resonance topologies. Method, data and licenses · Chart data

Resonance Routing — US 19/452,440.