Research program A
How models actually behave under realistic and adversarial task conditions, measured from observable input and output rather than from self-description.
Single-turn benchmarks report what a model can do once. Deployed systems are judged by what they do repeatedly, after context accumulates, after instructions compete, and after a newer version replaces the one that was tested. This program builds controlled tests for those conditions, so that a reliability claim names the conditions under which it holds.
The object of evaluation is observable behavior. A model's explanation of its own answer is treated as another output to evaluate, not as a view of its internal reasoning.
Do models keep constraints, corrections and earlier decisions intact across extended, multi-session work? This entry defines the question and the evidence that would answer it. It reports no results.
When a model explains a decision after making it, does the explanation match the evidence that was available to justify the decision? This entry defines the question without any claim about hidden internal reasoning. It reports no results.
