Local GateOne token. A decision.

Local Gate

Turn a local model into
yes, no, or a score.

I read one token's log-probabilities from a local Ollama model to give you a typed decision. No API key. Your data never leaves the machine.

Buy Local Gate, 19 USDHow it works

19 USD one time. Lifetime updates. 7-day technical refund.

The reliability dotsSix measured bins, raw confidence
The six raw reliability bins Traced from the supplied calibration chart. Observed accuracy by bin: 100, 75, 71, 83, 50 and 88 percent. Bin counts: 2, 4, 7, 6, 6 and 312. Positions reproduce the chart, not exact underlying mean-confidence data.

Measured, on my machine

Confidence needs calibration.

A raw 7B gate is overconfident. Temperature calibration fixes that on my measured decisions.

Calibration chart showing 337 labeled decisions, 87 percent agreement with truth, raw expected calibration error of 0.120, and six reliability bins.
My machine: qwen2.5:7b-instruct, RTX 5090 laptop. 337 labeled decisions; raw confidence agreed with truth 87 percent. Expected calibration error: 0.120 raw, 0.033 after temperature scaling, 0.040 cross-fitted. Warm decision: 30 to 250 ms.
0.120Raw calibration error
0.033After temperature scaling
0.040Cross-fitted calibration error

What you point it at

A small decision before the next step.

Lead triage

Filter leads before they page you.

Alert routing

Decide if an alert deserves a human.

Release gates

Gate a release as ship, review or block.

How it works

Install. Check. Serve. Ask.

  1. Install the tarball.

    Download it after checkout, then pip install localgate-0.1.0.tar.gz.

  2. Check the local setup.

    localgate doctor

  3. Start the local server.

    localgate serve

  4. Ask for a decision.

    Use localgate ask or POST /ask.

POST /ask
POST /outcome
GET  /dashboard

A real exchange from the author's machine. Choice questions return the winning option and the probability of every option. Number questions return a value from 0 to 1.

POST http://127.0.0.1:8890/ask
{"state": "Deploy failed twice tonight with a database timeout. Rollback is one command. No customers affected yet.",
 "role": "an on-call lead",
 "questions": {"act": {"instructions": "Roll back now or keep debugging?",
                       "options": {"rollback": "roll back now", "debug": "keep debugging"}},
               "urgency": {"type": "noul", "instructions": "How urgent is this?",
                           "low": "can wait until morning", "high": "wake the team up now"}}}
{"answers": {"act": {"id": "5536eb450c066d62", "choice": "rollback", "confidence": 0.95,
                      "probabilities": {"rollback": 0.95, "debug": 0.05}, "calibrated": true, "ms": 41.2},
              "urgency": {"id": "108b14059d360d4e", "noul": 0.71, "confidence": 0.67,
                          "probabilities": {"7": 0.67, "6": 0.21, "8": 0.12}, "ms": 30.1}},
 "model": "qwen2.5:7b-instruct"}

What ships

The gate, and a way to check it.

  • Local HTTP server

    POST /ask, POST /outcome and GET /dashboard.

  • CLI

    localgate ask, serve, doctor, log, calibrate, judge and activate.

  • Decision log

    An append-only log of decisions.

  • Temperature calibration

    Fitted on labeled decisions.

  • Judge

    Labels pending decisions with the claude and codex CLIs when both agree.

  • Dashboard

    The reliability chart and hand grading.

Requirements

Bring a local model.

  • Python 3.11+
  • Ollama
  • One 7B model, about 5 GB

About 100 ms warm on a laptop GPU. A few hundred ms on CPU.

Pricing

One purchase. A local gate.

Local Gate

19 USD, one time

Single-purchaser license.

  • Lifetime updates
  • 7-day technical refund
  • No redistribution
Buy Local Gate, 19 USD

Payment through Stripe Checkout. License and refund terms.

FAQ

Before you buy.

Does anything leave my machine?

Your situations, questions and decisions never leave the machine. They go to your local Ollama model and to a log file on your disk. Two things do go out: the purchase, which Stripe handles, and a license check. Activation sends your key and a machine hash to my license service, and a licensed copy repeats that check once a week. If the check cannot reach the service it keeps working for 14 days. An unlicensed copy runs for 7 days, then asks for a key. Details are in my privacy page.

Which models work?

You need one 7B model in Ollama, about 5 GB. My measurements use qwen2.5:7b-instruct. I have not supplied a broader model compatibility list here.

How is this different from a hosted decision API?

The decision runs locally, with no API key. I read one token's log-probabilities to return a typed yes/no/score decision. You can use the CLI or the local HTTP server.

Is the confidence real?

It comes from the model's log-probabilities, but the raw gate is overconfident. On my 337 labeled decisions, temperature scaling reduced expected calibration error from 0.120 to 0.033; cross-fitted error was 0.040. Those are measurements from my machine, not a promise about every decision.

Can I use it in my own programs?

Yes. Call the local server with POST /ask, record outcomes with POST /outcome, or use localgate ask. The license is for a single purchaser and does not allow redistribution.

What about refunds?

7-day technical refund: if Local Gate does not run on your machine and I cannot get it running within 7 days of purchase, I refund the full price and revoke the key. Refunds are not offered for a change of mind, because the product is source code you have already received. Ask through the contact form on noumenon-ai.com.