AI now writes the code, screens the candidates, and drafts the decisions. We build the instruments that measure how these systems actually behave, and keep the evidence that holds up when it counts.
Get in touch Current research ↓Language models carry latent decision dispositions: identical scenarios with one attribute swapped can produce different judgments, inconsistently across vendors and shifting silently between versions. In regulated contexts (insurance claims and underwriting, candidate screening, lending) those dispositions carry legal consequence, and no validated way to measure them exists.
Our work is early-stage measurement science: procedurally generated matched-pair instruments with established reliability and validity bounds, calibration against implanted dispositions in open models, drift detection for stochastic, version-unstable systems, and contamination resistance for public evaluation grammars. The deliverable is not a score; it is a methodology whose results survive adversarial scrutiny, and the evidence file that goes with it.
Terry Harmer has spent more than twenty years building measurement systems inside regulated enterprises, most recently as a continuous improvement executive at a Fortune 250 insurer. MSc, Management & Information Systems, Trinity College Dublin. His career specialty is the discipline this research formalizes: measurement whose results hold up in front of executives, auditors, and regulators.