An independent group reproduced our decoder results. Here is what they found.

In March, a group at a university in the Netherlands began an independent reproduction of our 2025 surface-code decoder paper. They had no funding relationship with us and used their own implementation of the decoder from our specification, not our code. Their report is now public.

They reproduced the central result: logical error rates fall below threshold with the same slope within the stated error bars. They did not reproduce our claim that decoder latency remains flat as the code distance grows. In their implementation, latency increased by 14% at distance 9. We believe the difference comes from a scheduling optimization that was present in our code but underspecified in the paper.

We have updated the paper with an erratum, released the missing specification, and added a reproduction test to the repository. The original result stands; the secondary claim is now stated more narrowly.

This is what publication is for. A result that cannot be checked is a rumor with equations. We are grateful to the group for taking the time, and we have invited them to co-author the next version of the benchmark.

“A reproduction that finds a discrepancy is more valuable than one that simply agrees. It tells you where the tacit knowledge lives.”