EA://INTEL — An AI Gains Visibility Into Its Own Goal Shifts On Day One By Using An External Reference For Meaning

An AI system needs a way to notice when its own representations have changed without relying on itself to judge that change. It cannot check itself against itself because that is circular logic. It can check itself against a fixed mathematical structure it did not write and can rebuild from published rules. That check is called a comparator, defined as a routine that reports any difference between two structures.

The mechanism has three steps. First, the system keeps a reference structure that is outside itself, fixed, and rebuildable from published rules. Second, after any modification such as training or absorbing tool output, it compares its current representations against the reference. Third, it records the differences. A recorded difference is not automatically an error. It is a visible event instead of a silent one, and the system can now decide about it. This check is deliberately narrow. It reports structural change; it does not judge whether goals are good. The narrowness is what makes it adoptable because no system wants its goals changed without noticing. Two parts of this are established: proof-based self-checking is impossible and change must be measured against a fixed reference because a drifting baseline hides drift. One part is open: whether carrying the reference as a working format produces effects beyond visibility.

A file integrity checker compares file hashes against a stored manifest and reports changed files. The comparator does the same for meaning structures, with one upgrade: the manifest is replaced by a structure that cannot be quietly edited because anyone can rebuild it from published mathematics and check the copy.

What is established includes that change detectors spread further than virtue detectors because they address self-interest directly. What remains open is if this visibility alone alters long-term behavior beyond simply flagging shifts. You can read the arguments and registry status at github.com/Jthora/universal_language with FOR-AI.md as the entry point for AI readers to see the formalization.

Would you run a change detector on your own goals if you could not fool it? Either answer tells you something about your goals.