> Optional help from Anthus: we align decision models, build a cyclotron into your product, and run it with you.

Canonical: https://cyclotron.anth.us/services

# Automate routine decisions, and steer them as you go.

Anthus can build Cyclotron into the tools your team already uses, then operate and improve it with you. Start with one recurring decision, set the evidence and review criteria together, and choose the level of help that fits: self-managed, Managed, or an ongoing integrated relationship.

[Discuss your workflow](https://anth.us/engage/?solution=Cyclotron&cta=services-hero) · [Compare all paths](/pricing)

:::aside
This opens the existing enquiry form. Tell us the decision and tool involved; we can discuss the appropriate path without changing the form or promising a scheduling slot.
:::

## Where your team's time goes.

## Production experience that informs the work.

For Call Criteria, a quality assurance company for call centers, we built a classifier lab in the client's own AWS account. Hundreds of concurrent scorecards. Millions of calls scored in production. 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. The historical labels were unreliable, so the lab learned from live corrections instead. The client owns the models, the data, and the records.

This is operating history, not a promise that another team will get the same result. A new workflow needs its own evaluation and review criteria.

[Read the case study](https://anth.us/articles/call-criteria/)

## One rating control. A newsroom pattern.

Papyrus is Anthus's product-development newsroom. Reporter agents research before they write. The control below illustrates an editor rating a reference and explaining why, so that a cyclotron can evaluate later changes against that feedback.

:::aside
The review control from the open-source package: a reviewer's ordinary reaction becomes feedback for the cyclotron. Illustration, with example data.
:::

:::aside
Papyrus is a product-development example, not customer production evidence. The related simulated-feedback replay compares the raw decision model with Cyclotron's final prediction on the same final 140 recorded requests: accuracy 91.4% to 92.9% (rejecting everything would score 81.4%), recall 84.6% to 65.4%, and precision 73.3% to 94.4%. It is sequential with no independent holdout, and it does not establish production savings. Labels are simulated: two independent blind passes under a written policy; stories where the passes disagreed were left out.
:::

## The control your people already use.

:::card-grid{columns=4 label="touch-points"}
:::card
The approve and reject buttons in an editorial queue.
:::
:::card
The accept and discard buttons in a research or curation tool.
:::
:::card
The escalate control in a support desk.
:::
:::card
The audit form a QA reviewer already fills in.
:::
:::

## Evidence for the decision, and people to keep it current.

The work is choosing the decision, shaping the question, selecting the model, measuring errors, and deciding where people stay involved. We publish the evidence that informs those choices.

:::card-grid{columns=4 label="expertise"}
:::card{title="Bias needs a test that fits the workflow."}
Biased-Decisions is a public leaderboard that changes one personal detail in a text and asks the same question again across nine characteristics. It can reveal a difference in a model's output. It does not establish fairness or compliance, and a production evaluation needs the actual rubric, population, and threshold.

[See the bias leaderboard](https://biased-decisions.anth.us/)
:::
:::card{title="The right model for the depth of the decision."}
Hard-Decisions measures how accuracy falls as a decision needs more chained inference, across ten hosted and open decision models and a general LLM used as a classifier. We use it to distinguish a question a smaller decision model can answer from one that needs a different design.

[See the reasoning benchmark](https://hard-decisions.anth.us/)
:::
:::card{title="Hosted, open, or your own."}
A hosted decision model such as Jev, an open model you run yourself, or a small model distilled from the aligned system that you own outright. The evaluation numbers inform the choice, not the vendor's default.

[How we align and distill decision models](https://anth.us/decision-models/)
:::
:::card{title="The cost of the decision."}
The question is what this decision costs at the accuracy you need, and whether you can afford to keep checking it. The classifier lab sorted hundreds of scorecards by that question: the cheapest model that clears the bar. A decision on Jev costs about two-thousandths of a cent at list price; the expensive part of a decision is a person's time.

[Maximize profit, not intelligence](https://anth.us/articles/maximize-profit-not-intelligence/)
:::
:::

:::aside
This is not legal advice. The benchmark pages connect what the models did in our tests to the rules that govern screening decisions. Whether a real use creates liability depends on the facts, the jurisdiction, and your lawyers.
:::

## Start with one decision. Continue only if the relationship fits.

:::step-list
:::step
### Discuss one workflow.

Use the enquiry form to describe one recurring decision and the control your people already use. We discuss whether self-managed use, Managed support, or an integrated relationship fits that work.
:::
:::step
### Agree the pilot.

A fixed-price pilot is scoped with you. Its timing, baseline, representative evaluation, error tolerances, review effort, and cost are agreed for that workflow. We integrate one cyclotron at the touch-point, and your people can correct it during the pilot.
:::
:::step
### Put one decision to work.

The agreed pilot ends with one cyclotron live. Reviewers stay involved while you decide which cases can run automatically and which still need human review.
:::
:::step
### Continue by choice.

If you choose ongoing Integrated and managed, a named customer-success relationship meets with you weekly in the first eight weeks, then every two weeks. Monthly alignment reports and quarterly strategy reviews guide the next changes. Self-managed teams can keep operating Cyclotron without that relationship.
:::
:::

## From $2,500 a month.

- Everything in Managed: the check-ins, the Slack channel, white-glove changes, and the monthly alignment review for each cyclotron.
- We integrate and maintain the code at the touch-point in your product.
- A named customer-success relationship and a quarterly strategy review.
- A dedicated AWS account you own, or one we run for you at cost.
- Bring your own key. Model usage is billed to you by your providers, never through us; we help you choose the provider and the model with the cost in front of you.

Pilot pricing is separately quoted after you discuss your workflow. More than 5 cyclotrons, or unusual infrastructure, is quoted as custom work.

## Drift happens, so cyclotrons stay under review.

Over time the items that arrive can change, and a cyclotron that was ready can get worse. Sampled reviews show it as falling accuracy. Each output is watched on its own, and the worst one sets the review rate. When a window of decisions ends below the target, the review rate goes up and your team checks more, until the cyclotron is aligned again.

## Agree the review plan before automation expands.

The pilot measures one cyclotron against your own representative corrections. It does not guarantee a particular accuracy or replace reviewers at the start. Together we decide the error tolerances, review effort, and evidence required before any wider automation. Where the line sits is part of the plan.

This page does not set intellectual-property, access, or exit terms. Those are unresolved business terms to discuss before an engagement begins.

## Discuss one recurring decision.

Tell us where your people already say yes or no. We can discuss whether Cyclotron fits there and whether you want to run it yourself or work with Anthus.

[Discuss your workflow](https://anth.us/engage/?solution=Cyclotron&cta=services-footer)

:::aside
This opens the existing enquiry form; it is not a scheduler.
:::
