Is it a fit?Book a call

How it learns

Keeps learning. Released only when better.

Your people approve the work. A new version trains on it and faces a sealed test. Four lights say what happened.

  1. LearningApproved work trains a candidate.
  2. ReleasedBetter on your sealed test. Live.
  3. HeldNot better. Nothing changes.
  4. Rolled backBack to the last good release.

In use

It learns the work you do.

Choose an expert. Review the work. Decide what it learns.

  1. Offers in house styleReleased
  2. Contract review
  3. Claims triage
  4. KYC review
  5. Sorting referrals
  6. Answers on local rules
  7. Operations triage
  8. Price forecasts

Illustrative: one model’s experts, each with the light of its latest release.

  1. 01ChooseA permitted expert for the task.
  2. 02AskWith approved sources.
  3. 03ReviewCorrect the draft.
  4. 04ApproveMark what it learns.
  5. 05TrainFrom 50 approved examples.
Shown as designed. Today a prototype runs on test data.

A correction is not training data until someone approves it.

A candidate is not a release until it passes the gate.

50

approved examples before a training run can start.

Learning

Approved work becomes a new candidate, in three steps.

  1. DistilA large open model answers your task prompts. Only answers that pass your graders are kept, and your expert learns from them.
  2. PractiseThe expert tries one real task several times. Your grader scores each try, and the expert learns from its best.
  3. GradeEvery candidate is scored the same way, with pinned data and seeds.
The scoring harness runs today. Learning in production is planned.

Released

A candidate beats the live version on your sealed test set and goes live. The record gains a line.

A woman at a wooden desk reviewing a model’s screen on her laptop
Keller TLM
Release record

  • Release 13

    Learning from approved work

    Learning

  • Release 12

    22 October 2026

    Released

  • Release 11

    8 October 2026

    Released

Illustrative data. The live version only changes when a release beats it.

Held

Not better, so nothing changes. The live version stays, and the candidate stays in the lab. A tie keeps the live version too.

Our own gate would hold Benchmark 001’s version: better than its base model, not yet ahead of a simple last value baseline.

Base model1.750%
Our first training run1.439%
Last value baselineStill ahead of ours1.392%
Our first training run cut forecast error 17.8% against its base model. It is not yet ahead of a simple last value baseline, and we publish that too. Benchmark 001 · forecast error (CRPS), lower is better · 243 held-out days, hourly BTC and ETH, 24 hours ahead · run on 14 October 2026.

Rolled back

Back to the previous release, at any time, by the owner. Rollback in 60 seconds is a design target.

60 s

Every release on record

The owner figure moves one step with each passed release.

Every model has a figure of its own, drawn by three numbers from the model: angle, pitch and curve. It sits on the Model Deed, the device and the app.

  1. Release 1
  2. Release 3
  3. Release 5
  4. Release 8
  5. Release 10
  6. Release 1215.5°, 1.035, 0.26
Example owner figure 3302, at releases 1 to 12. Illustrative. Under the latest: its angle, pitch and curve.

Start

See it on your own work.

One task, a sealed test, and a release only when it is better.

Book a callRead Benchmark 001