Skip to main content

Cyclotron

SELF-ALIGNING AI DECISIONS

OPEN SOURCE · MODEL-NEUTRAL

Automate routine decisions, and steer them as you go.

A cyclotron is one AI decision, such as publish or reject, that learns from your team’s corrections. A general-purpose AI does not know the details of your business. A cyclotron picks them up from the corrections your team already makes, so people steer the decisions instead of making every one themselves. A monitor shows when it is ready and how it holds up: how often its answers match your team’s labels, and whether it is as sure as it should be.

Run the open-source library yourself, or ask Anthus to manage a cyclotron or integrate one into the tools your team already uses.

Discuss your workflowRun it yourself, free

Support ticket triage

Steer

itemcyclotronhow-tohow-tobillingdecisionreviewerrecent decisionsaccuracyyour team's time

Your team corrects each miss, so it learns your standards.

Illustrative examples. The marks and gauges show direction, not measured results.

Calibrated confidence

Send the unsure ones to people, and know they are the right ones.

Calibration, in detail →

Sure, but wrong

most sureleast surepeople checkbelow this lineslipped throughcyclotronrevieweralready rightmistakes that slipped throughreviews spent on right answers

A general model's confidence does not match your data, so wrong answers can look sure and skip review.

Illustrative example. Positions show direction, not measured results.

A recorded example

What the recorded runs show.

Recorded on Wikinews stories with simulated editor feedback.

All recordings →

Accuracy climbs, from a blank rubric.

0%25%50%75%100%1–7071–140141–210211–280stories in each window
  • blank start
  • first rubric sooner
  • rejecting everything
With no starting rubric, accuracy went from 44.3% in the first window to 95.7% in the last; a setup that proposed its first rubric at story 21, not story 82, reached 77.1% in its first window. Each line is a single run on simulated editor labels, and each window holds different stories.

How much review it took.

0%25%50%75%100%rejecting everything: 84.3%91.4%11394.3%16795.7%280stories reviewed, of 280
  • 113 reviewed: lighter schedule from story 50
  • 167 reviewed: all for 50 stories, then half
  • 280 reviewed: every story reviewed
Accuracy in the last window of 70 stories, against how many of the 280 stories a reviewer labeled, for three review schedules from one blank start (filled circle: the cyclotron; dash: rejecting everything). One run per schedule on simulated editor labels, so a few points of difference is within the noise.

Since 2023, in production

Before Cyclotron had a name, we built this loop for Call Criteria, a quality assurance company for call centers.

In many instances, a new scorecard was set up with AI assistance within an hour before a critical meeting or high-pressure moment. That is initial scorecard setup. Validation and production rollout have their own evaluation and review steps.

Source: Call Criteria case study. “Scorecard” is Call Criteria's own term for a call-scoring rubric.

Hundreds

of concurrent scorecards in production

Millions

of calls scored in production

Within an hour

for many AI-assisted scorecard setups before high-pressure moments

Where a cyclotron fits

Any place your people already say yes or no.

All use cases →

  • Contact center quality
  • Editorial
  • Knowledge base
  • Support and trust

Work with us

Use the free library yourself, or choose help that fits.

Services →

Self-managed

Free

You host, operate, and align it. No account needed.
Managed

$500/mo

We keep it aligned in an AWS account you own.
Integrated and managed

From $2,500/mo

We build it into your product and run it.