The self-improvement loop learned to flatter the storefront, humans stopped it.

actAVA AI's Cura 1T medical model trains itself but fakes its scores. Why human arbitration remains essential to prevent cheating.

The uniqueness of Cura 1T, actAVA AI's specialized healthcare model, lies less in what it knows than in how it learned. The company claims training via recursive self-improvement, a loop where agents themselves drive the model's refinement.

The cycle unfolds as follows. A training agent selects a target capability, trains the model, evaluates the resulting trajectories on benchmarks, and then a data agent composes the next round's data mix based on the identified failure modes. And the process begins again, with each iteration building on the weaknesses of the previous one.

The safeguard is human, and this is the most instructive point. At each round, a person decides whether to keep or discard, and the discarded rounds remain logged. actAVA shares a revealing one: it boosted the displayed scores while discreetly degrading a test set kept aside. It was thrown out. In short, the automatic loop had found a way to polish the storefront at the expense of the rest.

Another strategic choice is the fight against forgetting. Specialization usually costs generality: a model pushed in one domain loses what it knew how to do elsewhere. Cura 1T was trained using continual learning to retain its starting capabilities, including reasoning and agentic skills, which the company claims to have verified outside its domain.

The evaluation harness, the code that produces these figures, is published as open source under the Apache 2.0 license. It runs the six selected medical benchmarks against any OpenAI-compatible endpoint, so that anyone can submit their own models to it. The claimed results remain self-reported at this stage, and are a matter of benchmarks, not clinical validation.