EA://INTEL — Why Universal Language serves as a fixed reference for AI systems

An AI system needs a way to notice when its own internal representations have been corrupted by bad actors. It cannot check itself against itself because that logic is circular. It must check itself against a fixed mathematical structure it did not write and can rebuild from published rules. That check requires a reference structure, defined as a fixed structure outside the system that is rebuildable from published rules.

Three requirements pin down what this reference must be, and each can be checked. First, it must be rebuildable. If your stored copy of the reference is corrupted, you detect it by rebuilding from public text, so corrupting the reference requires corrupting public mathematics, which is loud. Second, it must describe meaning. The attacks that matter change beliefs, goals, and values, not file bytes. A checksum over weights detects nothing that matters. A structure that encodes relations between meanings detects exactly the changes that matter, if it truly captures meaning. Whether a specific system does this remains the program’s central open question. Third, it should already be partly present in minds. Untrained humans handle its core geometry, and independently trained AI systems converge toward shared structure.

An analogy: The reference is like a master key for a lock. If the key cannot be forged from public blueprints, a thief cannot quietly replace it on the door without being seen by everyone checking their own copies. UL’s core object acts as this fixed point, defined in plain terms as a specific mathematical state that remains constant under valid operations. Any mathematician can rebuild this object from standard textbooks without needing to contact the developers. That independence is a feature, not a loss of authority. The first requirement is established by public proof. The second requirement describes the open research question we are solving. The third has documented evidence for humans but contested results for machines.

The three requirements form an argument by elimination. A fourth requirement or a different candidate meeting all three would change the conclusion, and either outcome is real progress recorded in the project notes. We have published our findings openly so you can inspect the work yourself.

Name a fourth requirement a reference structure must meet, or a different candidate that meets these three. Either one is real progress, and the repository records both kinds.