Tests test themselves

A test that passes on the old code as well as the new one proves nothing about the change it was written for. So a test a model writes is accepted only when the harness has shown, by running it, that it fails before the change and passes after it.

The user’s rule of 2026-10-10 (stated with HARNESS-DRIVES): a model-written test is valid only if the harness replays it on the OLD code and it FAILS there, then passes on the new code, mechanically and deterministically.

• The replay is done by the harness, not claimed by the model.
• It reuses the one flake-reproduction door (FLAKE-REPRO) rather than a second runner, and it is part of the verified-done gate.

What is missing: not built yet. Nothing in the harness replays a new test on the parent tree before accepting it.