OrbitCurve builds cybersecurity benchmarks, model and agent evaluations, expert trajectories, long horizon tasks, reinforcement learning environments, and focused datasets. Each project starts from a defined security capability or failure mode.

The scoring method is defined before the run. We use executable checks when the outcome can be tested directly and expert review when the task requires domain judgment. Capability, cost, latency, time, tokens, and compute are reported separately when they are relevant to the evaluation.

APEX CyberMicro Bench has a public repository. DeobfuscationBench is still in development. Customer specific evaluations, private task populations, and hidden checks are not published unless the engagement explicitly allows it.

Akula is OrbitCurve proprietary software for orchestrating cybersecurity evaluations. It connects the benchmark, model, agent, environment, execution policy, budget, grader, artifacts, and result record for each run.

Yes, when the deployment has been technically qualified for the customer environment. Network access, tools, data handling, infrastructure, evidence retention, and security constraints are agreed during scoping rather than assumed in advance.

Start with one security task or decision you need evidence for. We define the system under test, task population, success criteria, environment, execution policy, resource budget, and required outputs before any evaluation or data production begins.