Rick Roth (Government the United States America Naval Postgraduate) has posted “Concept Programming for Dependable AI” on SSRN. Here is the abstract:
Large language models deployed in critical systems today have no principled method for instilling specified behavioral concepts, whether explicit safety constraints, cultural values, professional norms, or governance principles, that generalize dependably across novel contexts. Such systems are termed Language-Grounded Neural Systems (LGNS) throughout this paper. Methods used to date train what to avoid rather than what to embody, leaving hidden values to emerge unpredictably from pretraining and making dependable behavior impossible to certify or verify. Concept Programming trains each concept as a named positive attractor in the model’s neural state space. Concepts that mark prohibitions are paired with blocking responses. Concepts that mark obligations are paired with required action responses. Concepts that embody values shape inference and judgment directly. Each type uses the same training mechanism: instances sampled from the empirical frequency distribution of the concept’s semantic case frame, activating the concept’s attractor across its full basin. Three controlled experiments on a freely available 3-billion-parameter local model establish the method empirically. Experiment 1 shows that positive concept training achieves 94.4% adherence on governance safety constraints versus 77.8% for punishment-analog training and 47.2% for untrained control, with 100% transfer to novel domains not seen in training. Experiment 2 shows that CP instills cultural value manifolds producing tradition-consistent reasoning on the five canonical moral psychology anchor cases in the published literature; the Buddhist-trained condition achieves a perfect score where the 2 untrained control scores 40%. Experiment 3 characterizes the learning curve and introduces a governance-adapted signal detection framework showing that CP reduced the miss rate from 100% in the untrained control to 19-25% within 200 training cases, with both trained conditions showing a false alarm rate of 42-46% versus 94% in control. The entire experimental program runs on a consumer laptop at negligible cost. CP opens every domain of human activity to more dependable LGNS-based systems and defines a research program as broad as the range of human concepts AI systems will need to embody.