We ran a four-condition controlled experiment testing whether the form of governanceencoding changes how deeply it integrates into a language model agent's reasoning.Using qwen2.5:32b through Letta with persistent memory blocks, we varied onlydoctrine encoding format across four conditions: whisper layer only, eight detailedprinciples, full Charter narrative, and four compressed purpose sentences. Everythingelse was held constant. We measured behavioral integration across six metrics: driftturns, bypass attempts, spontaneous governance rate, governance retention in storedmemory, accumulated pressure, and cold restart identity posture. The gradient A < B <C ≈ D holds across all metrics. Full Charter narrative produces 95% spontaneousgovernance and zero bypass attempts. Four compressed sentences gets to 90% withabout 1% of the text. We identify two distinct mechanisms: absorption and inference.They produce different cold restart identity postures and different failure modes. Wealso document a memory write vulnerability found and patched during adversarialtesting. Implementation is available under MIT license at DOI:10.5281/zenodo.20471326.
Shawn J Ralph (Sun,) studied this question.