QArena · Error correction Submit

QEC decoder accuracy on fixed circuit-level memory experiments

A decoder decides whether error correction works on real circuits. Fixed shots with hidden labels let decoders be compared on identical circuit-level data.

Rules

Scoring

Fewer logical errors over the 200,000 ranked shots ranks first; a shot counts once however many of its observables are wrong. Ties are ranked by submission number (lower first). Batch times are shown but never ranked. The verifier computes every score. The file format and the exact scoring rule are on the Overview tab.

Hidden labels and submission limits

  • Each account may have at most 16 scored or pending submissions per ranked instance, ever (accepted, queued or running files count; a rejected file returns no score and does not count). The daily limit of 30 files per problem also applies. Development instances have no lifetime limit.
  • Every submission to this problem stays private: no other account can download a prediction file, because a public file would let others probe the hidden labels through score differences. Only the scores are public. QArena's baseline files are public.
  • A submission file may be at most 1 MiB. Use one account per person; collecting scores through several accounts to reconstruct labels is not allowed.
  • Exact scores still carry information: one account's 16 scores can be worth some tens of errors per instance, up to about 100, through carefully chosen differences (weighing designs) or adaptive guessing. Differences of a few errors between rows are therefore not meaningful by themselves.

Batch times

Speed never affects acceptance or rank. A submission may report the time it took to produce the whole 200,000-shot batch (loading inputs, building the decoder, decoding, writing predictions) with its core count and hardware; it is always shown as an estimate. On request, QArena staff review a decoder program attached as decoder.tar.gz and time it on one CU Alpine CPU core (protocol named after the CPU model, for example alpine-7713-1c-v1): one untimed warm-up on 1,024 development shots, then three timed runs in fresh containers with no network and no access to labels; the median wall time is shown as verified only if all three outputs equal the submitted predictions byte for byte. A verified time says that QArena reproduced these predictions in that time on this fixed batch; it says nothing about unseen shots. Replacing the attached package hides the time.

Corrections

Records that depend on a verifier bug are removed once the bug is confirmed.

Access

Browsing is public. To submit, register with an email address and confirm it. Then upload files on a problem's Submit page, or create an API token on your account page.