Today's digest
How it works
The model reasons about its own failures, then directly rewrites the responsible weight values. No optimizer or training loop is involved.