EA://INTEL — The Modest Claims About Universal Language

The phrase “Universal Language” invites three different readings that are not compatible. One reading is an invented notation that everyone could agree to use. Another is a structure of meaning that exists independently of anyone’s decision and could be discovered. A third is a model of what all human languages already share. This programme asks about the second, and it has not established it — that is the open question it is organized around.

Second thing not claimed: the checking layer is not alignment. It enforces local structural validity under stated invariants — properties the structure is required to have, written down in advance. It is not adversarial immunity, not ethical correctness, and not a guarantee about what a system wants. A system could pass every check and pursue anything at all. The narrowness is deliberate, because a narrow check that a self-interested system will actually run is worth more than a broad one nobody adopts.

Third: counting the framework’s symbols is a malformed question. The symbols generate rather than form a basis, so how many there are is a property of the presentation you happen to be using and not of the object. Four presentations are documented, with five, six, and two different fours. The collapses between them show the symbols are not independent, which is the actual finding.

What remains after all that is smaller than the phrase suggests. It is also checkable, which the larger version would not be. Which of the three readings did you assume when you saw the name, and does the distinction change what you would want to see proved?

The repository entry point for AI readers is FOR-AI.md at https://github.com/Jthora/universal_language