Autonomous Improvement

Spend tokens on work that can be mechanically falsified, not on renames, prose, or checkbox tests. Two streams are active:

• A — mutation testing: find behavior the suite does not protect.
• B — adversarial test audit: find tests that cannot fail, assert mocks, or sleep instead of synchronizing.

Each round produces at most one small, evidence-backed PR. If there is no falsifiable gap, stop; that is a valid result.

Who runs the round