How do you control the way an AI system approaches a problem?
Our research is connected directly to how we build. Each finding answers three questions: What did we learn? Why does it matter for AI systems? What changes in practice?
Long-running indexes with a rubric, a score and an evidence trail.
Impersonato Cognitive Runtime: how a tested workflow is controlled in production.
Studies, case notes and methods — each tied to what changes in practice.
What we study.
What changes the work — and what only changes the wording.
What we learned. Generic and demographic-only personas sit on the fence 80–96% of the time. With explicit psychological structure, hedging drops to 26–46%, and the spread of distinct decisions widens 1.8×–4.1×.
Why it matters. A role label changes the wording more than the decision. Systems built on labels produce cautious, interchangeable output.
What changes in practice. We specify the procedures behind a perspective and keep only the parts that measurably change the decision, not just the words.
What we learned. Adding structure did not make output more synthetic: telltale AI artifacts fell from 12.9 to 3.3 per response, and flat “no real tension” answers dropped from 45% to 4%.
Why it matters. The common worry is that engineered behavior reads as artificial. Measured, the opposite held.
What changes in practice. We measure both task behavior and naturalness, instead of trading one for the other.
What we learned. Checked against a real advertising regulator’s own decisions, a committee of structured perspectives caught about 9 in 10 of the claims the regulator flagged — more than the strongest single model, especially on implied claims.
Why it matters. One pass from one perspective reads right past what a differently-oriented reviewer would catch.
What changes in practice. For review work, we compose several explicit review procedures rather than relying on one generalist pass.
Behavior is measured across Claude, GPT, Gemini, Grok and Llama model families. Each result is graded by confidence and checked against the published literature, with its limits stated alongside it.