Skip to content
← All news

News1 min read

17,520 QCode definitions across 21 lattices

QCode Discovery has searched 17,520 quantum-code definitions over 21 lattices and kept 5,178 that could win. Now nothing counts as a win without an exact construction, an independent check and a fresh replay of the IBM baselines. Humanfia-reported.

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. The gap between them is exactly where a long unattended run will settle if you let it.

The QCode Discovery pipeline now sits on the other side of that gap. It is fail-closed: a candidate becomes a result because it was certified, never because of its heuristic score.These figures come from the run's own records and have not been published anywhere else, so we label them Humanfia-reported.

Definitions processed

17,520

Every candidate code definition the search has scored so far.

Lattices

21

The lattices those definitions span.

Winner-capable, kept

5,178

About three in ten: definitions persisted because they could still beat a baseline.

A winner-capable definition is not a win. It is a candidate that the cheap metric says is worth the expensive check. Everything after that point is the part that changed.

What a win now requires ​

Fail-closed · 3 gates · 1 way out
candidates1construct2verify3replay IBMWIN■ refused■ certified

Swipe the diagram sideways to see all of it →

1/0 A winner-capable candidate arrives with a good heuristic score. The score alone counts for nothing.

An exact construction.

The code itself, written out, not a score that implies one exists.

An independent verification.

The construction is checked again, by something other than the process that found it.

A strict final replay.

The established IBM baselines the candidate is compared against are revalidated, in the same run, before the comparison is allowed to stand.

The third gate is the one that gets skipped in practice, and it catches the failure mode we care about most. A baseline drifts, or was measured under conditions that no longer hold, and a comparison looks like a win when the only thing that moved was the reference.

The run is allowed to be wrong. It is not allowed to be wrong quietly.

This is the same discipline as everything else here, and the same one the KDA² hardening arrived at for kernels: an agent optimizes the score it is given, so the score has to be the certificate.

Read next

News0% + 0.5% → 3.5% on ProgramBenchAlone, Opus-4.8 fully resolves 0% of ProgramBench and GPT-5.5 resolves 0.5%. As a builder and a reviewer in one loop they resolve 3.5%, close to the best official entry's 4.5%, and the curve was still rising when the four-hour budget ran out. A Humanfia-reported checkpoint.Read it BlogModel, Tool, Flow: 2, 18, 50 of 50The same model, measured three ways (raw API, the vendor's CLI, and that CLI inside a flow) on PutnamBench, Physics Cup, SuperChem and an HLE subset. Putting the CLI inside a flow adds more than the CLI's own tools do.Read it News#1 on Lean-Eval v1 first solvesLean-Eval froze its first 128-problem release and moved everything older to an archive. On v1 our GPT-5.6 flow has the most first solves, 30, and is one behind the leader on total; on the archive it has 170 of 171.Read it