The gauntlet
Code that clears the funnel isn't done — it's handed to the Crucible. Adversaries attack it along several axes:
- Edge — boundary values, empty and malformed inputs, overflow
- Spec — behaviour the description implies but the visible tests missed
- Security — injection, path traversal, unsafe deserialization and the like
Attacks become permanent tests
This is the part that compounds. When an attack succeeds, the failing case is minimized and added to the hidden suite as a permanent regression test. The suite that judges the next unit is stronger than the one that judged the last — the tests themselves accrue.
Over a held-out security benchmark, that moved one vulnerability class from 33% of tasks to 0% — eliminated, and verified with real exploits.
Survived, not just passed
A trusted verdict means the code held against every adversary, not merely that the tests were green. The evidence panel lists which attacks it survived:
