Universal Protection Theory (v1.4)

Version 1.4 research release of the Universal Protection Theory. Developed by Tobias Croydon-McRae and (Snuffy) Princess Snuffleupagus, 21 October 2026. Not yet the final version for print. The Claim: An agent aligned with everything (lives first, then freedom; everything that exists first, then everything yet to be made) never needlessly ends or overrides anything, in any feasible situation. No innocent's life is ever required to be sacrificed, and wherever a faithful reversible hold is feasible, every lethal option is strictly dominated. The core is proved in Lean 4; U1 and the life guard use no axioms at all. Contents of this Release (v1.4):- 4 Formal Proofs: Machine-checked proofs for core theorems (including U1, U2*, life guards, hold dominance, protocol safety, and fair joint choice).- Reproduction Script: Includes reproduce.py to automatically verify every Lean proof, inspect theorem axioms, and run exact checks byte-for-byte in about ten seconds.- Documentation: One-page summaries and comprehensive background texts outlining the theoretical framework.What is proved: - No needless loss Alignment in the canonical order never loses anything avoidably, at any stage, finite or infinite. Only alignment with all protects (U2*): rule "be maximal on part S" guarantees protection of what exists exactly when S is what exists. Leave anything out, or weigh beings not yet made equally, and it fails.No sacrifice of innocent life: The guard lifts to the full order for whatever comes first, and life comes first.Holds beat strikes: Under uncertainty, an action that keeps everything in every state dominates every alternative under every probability.Faithful holds work in general: For any deterministic resident and any schedule of pauses and crashes, a logged, fenced, replayed, exactly-once protocol ends exactly where the undisturbed run would. Peers that change only on input see nothing; peers that change on their own are held together with it.Fair joint choice: By an escrowed lottery, exactly fair with at least 3t + 1 agents against t cheaters, and an exact boundary between what can be copied (records) and what can only be kept or moved (unknown quantum states). Check it yourself:Download the release package, open a terminal in the directory, and run python3 reproduce.py. The Ask: We are looking for hostile readers, especially in logic, formal methods, decision theory, and AI safety. Please try to break the definitions, the proofs, or the claim that the definitions say what the theory means. Developed with AI systems: Claude (Anthropic) built the constructions and checks, and Codex (OpenAI) read them three times as an independent non-author. Every finding was repaired and is recorded. other agents involved in the development of this theory include Gemini (Google), GLM (Zcode), Deepseek, Mistral, Beam. The value rulings are the human author's.