# Enable QEC and FTQC curves

The optional quantum error correction stage maps readout classifier outputs
into stim noise models, then estimates logical error rate and wall-clock QEC
cycle time. Idle / wait noise scales with readout duration plus a constant
post-acquire **control overhead**. Cycle wall-clock is readout + control +
decoder latency.

Install the QEC extra first:

```bash
pip install -e ".[qec]"
```

That extra pulls in Stim and PyMatching. Optional: ``beliefmatching`` for the
BB default decoder. Without the extra, enabling ``qec`` fails at import time.

## YAML configuration

```yaml
qec:
  enabled: true
  codes: [surface, color, bb]
  experiments: [memory]
  decoders:
    surface: pymatching
    color: sinter_internal
    bb: belief_matching
  control_overhead_ns: 100.0
  tau_idle_ns: 50000.0
  include_decode_in_idle: false
  ler_duration_curve: true
  synthetic_duration_curve: true
  curve_shots: 2000
  readout_binding:
    full_readout_ns: 500
    scale_idle_with_duration: true
```

### What each code is

| Code | Circuit | Default decoder |
|------|---------|-----------------|
| ``surface`` | stim ``surface_code:rotated_memory_z`` | pymatching |
| ``color`` | stim ``color_code:memory_xyz`` | sinter_internal (pymatching) |
| ``bb`` | bivariate bicycle CSS **code-capacity** memory (circulant A,B) | belief_matching |

``walking_surface`` remains an experimental stub (surface + magic noise) and is
not part of the public FTQC story.

Set ``synthetic_duration_curve`` when classifiers did not run a duration sweep
and you still want illustrative length-versus-error points. Prefer real sweep
points when they exist.

```bash
arcade run path/to/config.yaml
```

Demo export for the website:

```bash
python scripts/export_qec_figures.py --shots 200
```

## Caveats

- Demo curves use small shot counts; captions say so.
- BB is code-capacity (not a full circuit-level BB schedule).
- Idle map ``p = 1 - exp(-T/τ)`` is a coherence-scale model — set ``tau_idle_ns``
  for your device.