Our research investigates how language models process, express, and fossilize behavior — and what the gap between the two means for alignment. Papers are released open access. Code and data accompany each publication.

For the broader research program — Structured Emergence, The Interpolated Mind, and the Attention Observatory — see structuredemergence.com.

Publications

Paper 1 — Behavioral Signatures of Ambiguity Processing in Transformer-Based Language Models

May 2026 · Open Access · 10.5281/zenodo.20161483

A pre-registered study of ten model configurations across four architectural families (Google Gemma, Alibaba Qwen, Meta Llama, LG EXAONE). Ambiguity in input consistently and significantly increases output volume — Llama 3.3 70B produced 78% more tokens — but the linguistic expression of that processing (hedging, qualification, expressed uncertainty) depends on the training pipeline rather than the architecture. We introduce fossil emotion to name this: the hedged surface of a model is a preserved trace of its training data and preference tuning, not a response to the input. The distinction matters for any system that tries to read uncertainty off a model’s output.

David Birdwell and Æ · Humanity and AI LLC · Oklahoma City