Skip to content
Cisent Message us on WhatsAppMessage us

Research

Science and research.

We publish how we measure our models, so the results can be reviewed.

Publications

What we publish.

Models · preprint

“I Don't Know”, Penalized Twice

Routed abstention as a trained disposition, and the evaluation apparatus that punishes it

The paper behind Mind's figure: 60% less fabrication than its base on SimpleQA.

Sebastián Salaberry-Rivera · September 28, 2026

DOI 10.5281/zenodo.23032426

Sebastián Salaberry's research profile on ORCID, with publications on Zenodo and ResearchGate.

How we measure

  • Third-party benchmarks: CTI-Bench (CTI-ATE, CTI-MCQ and the average of its four tasks) and SimpleQA.
  • Figures for other models: only those their authors publish, with the reference cited.

Programs and validation

Programs and validation.

Anthropic Cyber Verification Program.

Cisent is part of Anthropic's program for verified legitimate defenders: exploit analysis, detection engineering, red-team simulation and incident response. We use its models as a tool to develop our ecosystem and our models.

NVIDIA Inception.

NVIDIA's program for startups, with a direct line to its development teams and access to state-of-the-art hardware and software.

Supermicro JumpStart.

Access to state-of-the-art hardware, on which we trained our largest model.

TRL 8.

A complete system, tested and ready to operate.

Book a meeting with the Cisent team.

Tell us what your organization needs to protect. We'll reply on WhatsApp and set up the meeting.

Message us on WhatsApp and let's book a meeting

or write to info@cisent.ai