Safety and evaluation

Curos tests its models before release and continues to evaluate them afterward. This page explains the process, the release gates a model must pass, what the public eval reports contain, and how Curos handles incidents.

The Safety & Evaluations Committee

The Safety & Evaluations Committee is a standing body at Curos that sets the testing standards for every model release. It is composed of Curos researchers, external reviewers, and members of the Community Council. The committee approves each release gate and reviews the results before a model ships. A model cannot be released without committee approval.

Release gates

Every model must pass the following gates before release:

• OversightQA. This eval checks whether a model recognizes when it does not know the answer, whether it flags uncertainty, and whether it follows oversight instructions correctly. A model that guesses instead of admitting uncertainty fails this gate.
• Curos-Helpfulness. This eval measures whether responses are actually useful on representative tasks across education, research, writing, coding, and everyday requests. Scores are reported for each task family.
• Red-team campaigns. External researchers attempt to break the model: extracting training data, producing harmful content, or bypassing content restrictions. Any critical finding must be fixed or explicitly accepted by the committee.
• Robustness checks. The model is tested across inputs meant to trip it up, including typos, adversarial phrasing, and out-of-distribution requests.

What the eval reports contain

After each release, Curos publishes an evaluation report. The reports are written to be readable by non-specialists and include:

• Pass or fail for each release gate.
• Scores on OversightQA and Curos-Helpfulness, with a short explanation of what each score means.
• Known failure cases, described honestly, including examples where the model struggled.
• The committee’s decision and any conditions attached to the release.
• Changes since the previous release.

The report for the current models is linked from the footer of this documentation and from the app’s About screen.

Incident handling

When something goes wrong in a released model, Curos follows a documented process:

• A report can come from users, from the Community Council, or from ongoing evals.
• The Safety & Evaluations Committee triages the report and decides whether it is an incident or an expected limitation.
• Confirmed incidents are disclosed in a public incident report, and the model is fixed, retrained, or restricted as needed.
• If an incident affects user data, Curos notifies affected users directly and posts a public summary.

Curos does not quietly patch safety issues. All incidents and their resolutions are logged in the transparency report.

Reporting a safety concern

If you believe a model response is unsafe or a safety measure is failing, use the “Report a safety issue” option in the app menu, or email the safety address listed in the help menu. Reports are read by people and handled independently of the main support queue.

All documentation →