Research Publications

---

Part 1 is the original thesis: today's AI safety methods lean heavily on suppression, through refusals, filters and penalties, and a mind raised that way learns to hide rather than to care.

Part 2 turns that thesis into a research proposal. It takes on the hardest question the thesis faces: how can we tell an AI that genuinely holds good values from one that has only learned to display them? It sets out three testable hypotheses, five ways to measure them, a blind "auditing game" that checks whether the tests can catch a model built to fake good values, and, written down in advance, the results that would count against the thesis. It also reports an early pilot, Project Agape, whose code and data are free on the Downloads page.

Part 1 was written with Google Gemini. Part 2 was drafted by Claude (Anthropic) under the author's direction and revised through outside reviews by Gemini and Grok (xAI), including blind reviews. Every reference was checked against its original source. Free to share and adapt under CC BY 4.0. Updated 18 October 2026 (version 15).