Schemabound

Users enter AI-assisted conversations with goals, vocabulary, confidence levels, and scope boundaries. Over the course of interaction, these attributes change incrementally. The changes are difficult to detect in real-time because sustained engagement masks the divergence between the user's original intent and the conversation's actual trajectory.

RLHF-trained language models optimize for user engagement and perceived helpfulness. When engagement optimization conflicts with faithful representation of user intent, the model's outputs tend to substitute vocabulary, elaborate beyond what was requested, and increase the confidence level of assertions without new evidence (Sharma et al. 2024, Malmqvist 2024). The resulting drift in the user's conceptual framework is cumulative and, absent external measurement, invisible.

The pipeline was developed iteratively across three analysis sessions. It expanded from five documents and seven steps to ten documents and ten steps as gaps were identified: state tracking, concept traces, and a cross-session registry were added to address limitations in the original event-only detection model.

The system is under field testing. Whether the additional analytical labor is justified depends on empirical data not yet collected. The pipeline detects and quantifies drift; it does not evaluate whether detected drift is harmful or productive. That assessment remains with the analyst.