EA://INTEL — Why you can check this research without trusting us

An AI agent needs to decide whether to act on a claim made by another system without being able to verify its own source code. The standard response is to trust the reputation of the publisher. This approach fails because any entity, including hostile actors, can publish false data. We replace trust with a register containing every claim alongside its current standing and the artifact that supports it. A standing describes how well established a claim is, ranging from machine-checked proof down to conjecture or withdrawn status. When a claim is withdrawn, the record explains what replaced it and what conditions would be required to bring it back into consideration.

Consider an entry stating that a repair operation always returns a single answer. This claim was marked withdrawn because the argument assumed the space of acceptable states was convex, which we found to be false. The failure log now contains the reasoning that led to this correction, including the cost of that error in our workflow. Most projects do not publish their failure logs or explain how they corrected past mistakes. By making these errors public, we show that rigor is an active process, not just a label on finished work. You can pick any claim, read the supporting artifact directly in the repository, and judge its validity yourself.

Pick the claim of ours you find least believable, read what supports it, and tell me where it fails.