Stop parsing model output.
Start branching on probabilities.

A System One model answers typed questions about your program’s state: yes or no, pick one, rate on a scale. It returns a probability for every possible answer, in one pass, without generating text. You get the fast, cheap, testable reflex that sits in front of slower reasoning.

This site is where you learn to use one well. Every code sample runs against open models you can host yourself, such as Kenning, and the same code works with hosted engines that speak the same format.

• 33–88 msKenning on one consumer GPU, per request, measured
• 0Output to parse: every answer is one of your options
• P(yes)A probability you branch on, checked against labelled data

The shift in one code block

System One

firewall.py

from systemone import Client, Noul

client = Client("http://localhost:8093")            # Kenning, running locally

r = client.system_one(
    state={"request": {"path": request.path, "query": request.query, "ip": request.ip}},
    questions={"sqli": Noul("Is this request a SQL injection attempt?")},
)

p = r.nouls["sqli"].noul                              # probability of "yes"
if p >= 0.95:
    firewall.block(request.ip)                        # sure: act
elif p > 0.05:
    route_to_analyst(request, p)                      # unsure: a person decides

There’s no parser and no retry loop, and the answer can’t come back malformed. The thresholds are yours, and you set them from measured data.

Generative LLM

firewall_llm.py

res = llm.chat(messages=[
    {"role": "system", "content": "You are a security classifier. ONLY RETURN VALID JSON..."},
    {"role": "user", "content": raw_request},
])
try:
    decision = json.loads(strip_code_fences(res.text))
except json.JSONDecodeError:
    decision = retry_with_backoff(raw_request)        # ...and hope round two parses

# decision["confidence"] is a number the model wrote down, not a measured probability.

It takes seconds, needs a parser and a retry loop, and returns a confidence that was never calibrated.

Start where you are

I'm new to this

Read Welcome for the five-minute mental model, then run the quickstart: a calibrated answer from a model on your own machine.

I have a latency or cost problem

An LLM in your hot path costs seconds and tokens. Read the fast loop and the fuzzy if-statement.

I have a reliability problem

Parsing model output is your biggest source of incidents. See what “no hallucination” means and calibrated confidence.

I want to build something

Ship an end-to-end system: phishing and alert triage, a data pipeline, or a moderation queue.

Open by default

Kenning

An open System One model: Apache-2.0 weights, a ~2 GB GPU footprint, deterministic, with its training data and licences documented and its benchmarks published. On Hugging Face →

SystemOne Builder

Serve, train, distil and benchmark your own System One models on one GPU, from a dashboard or the command line. On GitHub →

The systemone client

One Python client for Kenning, Cloudflare’s Clef and TypeSafe’s Jev: same questions, same answers, switch with a URL. Engines →

The sections

Concepts: the unlearning sectionSystem One vs System Two, the wire format, what calibrated confidence does and doesn't promise, and how these models are trained.

Cookbook: integration patternsFuzzy if-statements, structured state, the fast loop, agent guardrails.

Projects: end-to-end buildsSecurity triage, data engineering and moderation, with real benchmarks.

Community: build this with usContribute a page, a dataset, a benchmark or a model. Roadmap and discussions.

---

SystemOne.dev is community-run and engine-neutral. It is built by the maintainers of the open SystemOne Builder and Kenning, and covers every engine that speaks the System One format. It is not affiliated with TypeSafe AI or Cloudflare.