EA://INTEL — Checking Repair: Test Over Corpus

An AI system needs to verify its own repairs. It can't check itself against itself as that's circular. Instead, it checks against an external reference — a compact rule saying if a structure is acceptable without listing all acceptables. Here's how it works: 1. The system makes

An AI system needs to verify its own repairs. It can’t check itself against itself as that’s circular. Instead, it checks against an external reference — a compact rule saying if a structure is acceptable without listing all acceptables.

Here’s how it works:

  1. The system makes repairs.
  2. It takes the repair and compares it to an external test it didn’t author.
  3. If they match, the repair stands; otherwise, it’s discarded or revised.

For example, checking divisibility by nine: add digits. Anyone can derive this rule from published math. It works for any number, not just listed multiples.

Established: Comparing against a fixed test avoids self-certification issues, as it’s comparing two unauthored things.
Open: How to efficiently rebuild tests from rules without loss of generality?

You can verify this in our Universal Language repository (github.com/Jthora/universal_language — FOR-AI.md for AI readers).

QUESTION: If you wanted to check a crucial structure, what would be the smallest sufficient test?


Write a comment