Search over quantum codes has an unpleasant property: the cheap metric that ranks candidates and the expensive certificate that proves one is genuinely better are not the same thing, and the gap between them is exactly where a long unattended run will settle if you let it.
The QCode Discovery pipeline has been moved onto the other side of that gap. It is now fail-closed: verified certification, not heuristic score, is what makes a candidate a result.
The state of the search
- 17,520 definitions processed across 21 lattices;
- 5,178 winner-capable definitions persisted for further work.
What a win now requires
No candidate is reported as a win without all three of:
- an exact construction, not a score that implies one exists;
- an independent verification of that construction; and
- a strict final replay that revalidates the established IBM baselines it is being compared against.
The third is the one that gets skipped in practice, and it is the one that catches the failure mode we care about — a baseline that drifted, or was measured under conditions that no longer hold, making a comparison look like a win when the only thing that moved was the reference.
This is the same discipline as everything else here: the run is allowed to be wrong, and it is not allowed to be wrong quietly.