All guides

Understand

What Cyclotron does

When your team makes the same kind of judgment repeatedly, its corrections contain useful information. Cyclotron gives you a way to use that feedback, test proposed improvements, and keep a record of how the decision changes.

Last reviewed:

Start with one recurring decision

A decision model answers a defined question about an item. Should this source enter the library? Does this call meet a particular standard? Should this ticket be escalated?

Cyclotron works around that decision. A model can give an initial answer before you have collected a training set. Your reviewers then supply the task-specific evidence needed to improve and evaluate it. That starting point is useful for trying a task; it does not establish that the initial answers are ready to act on without review.

What happens in the loop

A correction supplies evidence. It does not guarantee that every metric improves, or that the next answer will be right. The technical explanation shows the parts of the loop and how evaluation fits around them.

  • The system makes a decision and records its confidence.
  • A reviewer checks the result, corrects it when needed, and can explain why.
  • Feedback helps fit a small learned model. An LLM optimizer can also propose changes to the written rules, examples, or questions.
  • Proposed versions are evaluated before promotion. Cyclotron records decisions, feedback, versions, and evaluation evidence.

A simple example

Imagine a team selecting sources for a knowledge library. People already accept useful sources and discard unsuitable ones. Those choices can provide feedback for a decision about what belongs in the library.

The team still needs to define useful, check disagreements, and measure errors on representative examples. Its integration determines how model answers enter the existing workflow and which actions require a person.

Is your task a good fit?

A precise-looking probability is not enough. The decision-model probability experiment shows why returned probabilities need an outcome check.

  • Does the same kind of decision recur often enough to evaluate?
  • Can reviewers explain the standard and resolve disagreements?
  • Can you collect representative examples, including difficult cases?
  • Do you know the consequences of a wrong answer?
  • Can you retain review or restrict automation while you learn?

Choose how much help you need

Cyclotron is open source and can be run self-managed. Anthus also offers Managed support and an ongoing integrated option for teams that want help operating and improving a cyclotron in their product.

Whichever path you choose, begin with one decision and agree what evidence and review are needed before wider automation.

Supporting reading

Inspect the research and technical explanation.

Next

Choose a practical next step.

Continue learning

The other practical guides.