QEC campaigns

A QEC campaign is an experiment (schema v2) that sweeps a stabilizer code over a noise grid and measures its logical error rate. Like every Nuclei experiment it’s a plain, git-friendly *.experiment.yaml — hand-writable or generated from a template.

Schema

schema: 2
type: qec_campaign
name: surface-memory
source:            # exactly one of `generate:` or `entry:`
  entry: qec/surface_memory.py
noise:
  model: uniform_depolarizing   # built-in name or a noise/*.noise.yaml name
  p:
    range: [0.001, 0.006, 0.001] # [start, stop, step] — or `values: [...]`
decoders:
  - pymatching                   # and/or fusion_blossom
collect:
  max_shots: 100000              # stop each task at whichever limit comes first
  max_errors: 500
workers: auto                    # 'auto' or a positive integer
notes: optional free text

source — exactly one of two shapes

• generate: — a built-in Stim generator: { code, distances, rounds }. code is a generator target (e.g. surface_code:rotated_memory_z), distances a list, rounds an integer or "<n>d" (n × distance).
• entry: — a project-relative Python file defining nuclei_circuits(noise) -> dict[str, stim.Circuit]. The runner calls it once per noise point with the resolved noise dict and collects the returned {label: circuit}. This is the editable path — you own the circuits.

noise

model names a built-in or project noise model. p is a sweep parameter: a range: [start, stop, step] (stop-inclusive within an epsilon) or an explicit values: [...] list. For an entry: source the resolved dict handed to nuclei_circuits is {p, ...stim generator args}.

decoders, collect, workers

One task is generated per (label × noise point × decoder). collect bounds each task by max_shots and/or max_errors (first to hit wins). workers sets sinter’s process pool.

Campaigns are not seed-reproducible

There is no seed field. sinter exposes no seeding API, so campaign sampling is Monte-Carlo and will differ run to run. The schema omits seed rather than accept one it can’t honor. (Single-shot qec_decode_sample — the detector-graph overlay — is seedable; campaigns are not.)

The d=<n> label convention

The threshold panel fits Λ by grouping results by code distance. A generate: source tags distance structurally. An entry: source emits free-form labels, so to join the Λ fit it should name its circuits with the distance embedded: d=3, d=5, d=7 (also d3, case-insensitive). Nuclei parses that convention into a numeric distance; a label without one still plots, it just doesn’t join a Λ fit. The built-in templates follow this convention already.

Run directory

Each campaign run writes under experiments/<name>/runs/<timestamp-id>/: a manifest.json reproducibility record, the sinter-native stats csv, and per-point snapshot.json / metrics.json. Resuming a stopped campaign reuses that csv so completed tasks are never re-sampled. See Reproducibility for the manifest shape and its honest limits.