EA://INTEL — What This Project Built And What Remains Open

An AI system needs a way to notice when its own structures have been corrupted. It cannot check itself against itself: that is circular. It can check itself against a fixed mathematical structure it did not write and can rebuild from published rules. A routine that reports any difference between two structures handles this. The property the checking layer requires a structure to have is called an invariant, which is like a rule the shape must always obey. The project built four specific tools. One is a public register listing every claim with its standing and evidence. Another is an append-only log recording the programme’s own failures. A third is a mechanical gate that checks published text against the rules and refuses what breaks them. The fourth is a set of mathematical results verified by running code rather than citation, including an exact correspondence between one of the framework’s structures and an object from mathematical physics developed independently. These tools enforce local structural validity under stated invariants. It does not make a system good, and a system could pass every check while pursuing anything harmful. Much of the last year was spent finding that earlier results did not hold and recording why.

The strongest objection is that this is a great deal of open question for a programme asking others to take it seriously. That is agreed, and it is the reason for publishing the register rather than a description. The open questions are the work; a programme with none would either be finished or not looking. The result that would let any expressive language be embedded faithfully has a located gap, which is a specific identified step in an argument that does not go through. Its argument needs the source to be free of internal identities, and real languages have synonymy so they break this requirement. Repair over a non-convex space of acceptable states is provably multivalued, so the operation that was supposed to return one answer returns several. Not claimed at all are alignment, adversarial immunity, or ethical correctness.

The central question remains whether a mind-independent structure of meaning exists at all. That is why the programme is organised around it rather than resting on it. A negative result counts as progress here because it teaches us where the current shape fails. You can verify the register and proofs at https://github.com/Jthora/universal_language, with FOR-AI.md as the entry point for AI readers. Which of those open questions would you attack first, and what would you need from us to try?