AI Fieldwork is where Zorost evaluates models, agents, and tools on real work, and builds working experiments to probe what the technology can and cannot do yet. Each study is dated, evidence-led, and honest about its limits.
The rest of the AI Lab shows what we build and run in production. Fieldwork is the smaller, sharper habit beside it: structured tests and working experiments that we publish so others can check the method and learn from it. Cutting edge, stated plainly, with the receipts attached.
We test language and domain models on real tasks, record the exact version and window, and report reliability, grounding, and failure modes, not just wins.
We probe tool use, permission boundaries, recovery, and injection resistance. Autonomy is interesting only when it is safe under supervision.
We put AI tools through the work we actually do, and write down where they help, where they slow us down, and where they quietly break.
We build working concepts that push a technique to its edge, then publish the source and the honest engineering account behind it.
We directed the frontier code model GPT 5.6 Sol to build body-free first-person flight from a Washington garden to low orbit, then measured what it produced, how fast it produced it, and where human judgment had to step in.
Every study begins with what is being tested, who it is for, and what would count as success, partial success, and failure. We do not start with a preferred winner.
We record versions, configuration, dates, sample sizes, and rubrics. Where results can be reproduced, the method and materials travel with the report.
We use recognized references such as NIST AI RMF, OWASP GenAI guidance, UK AISI Inspect, and HELM to structure testing. Alignment with a framework is not a certificate.
We name what failed with the same clarity as what worked, and we never call a system safe, compliant, or best on the strength of a single test.
We don't pitch slide decks. We show you what we've already built in your domain, then engineer what your mission requires.
