01Published work
Methods, evidence and limits.
01
Grounded execution / September 2026
Programming with Classifiers
The worked example, Lecter, retrieves exact source evidence. It is a small, inspectable operation with a replayable companion, not evidence of a general reasoning system.
Exploratory experiments on small, partly reused cohorts. Internally reviewed with separate agent checks; not externally peer reviewed. Replay verifies execution consistency, not fresh model performance.
02
Representation robustness / September 2026
Same facts, different decisions
How much should a decision change when the facts stay the same but their representation changes? This investigation examines representation robustness in Jev.
Results apply within the tested conditions. Internal review used independent code implementations and coding-agent assistance; the study has not been externally peer reviewed.