Pipeline stages

When you invoke arcade run with a YAML configuration file, ARCADE executes a fixed sequence of stages. Understanding that sequence makes log messages and output directories predictable. It also clarifies which YAML sections must be present for a minimal run and which sections only matter when you opt into optimization or quantum error correction.

        flowchart LR
  A[Data] --> B[Features]
  B --> C[Classify]
  C --> D[Hardware]
  D --> E[Report]
  C -.-> O[Optimization]
  C -.-> Q[QEC]
  O -.-> D
  Q -.-> E

DataLoad, demodulate, split, cache
FeaturesPer FEATURE_KIND matrix
Classifyfit / predict for each method
OptimizePrune, distill, NAS, tune
HardwareAnalytical LUT / DSP / latency
QECLER and cycle-time curves
Reportsummary_report + plots

Data preparation

The data stage loads raw in-phase and quadrature traces from HDF5 or from other supported loaders, applies demodulation and preprocessing according to the recipe in the configuration, splits shots into train validation and test sets with a frozen seed, and materializes caches so later stages do not repeat expensive work. Feature extractors that are shared across classifiers may also run here depending on configuration. The result of this stage is a prepared bundle of arrays and metadata that every classifier in the run will share.

Fairness depends on this stage. If two classifiers saw different demodulation settings or different splits, their accuracy numbers would not be comparable. ARCADE therefore prepares once and classifies many times.

Classification

The classify stage iterates over the names listed under classifiers.run. For each name it constructs the registered class, fits on the training portion of the prepared data, evaluates on the held-out portions, and records metrics into a structured result object. Classical methods such as thresholding and linear discriminant analysis tend to be fast. Neural methods may train for many epochs and write checkpoints. Hybrid methods may expand features first and then fit a linear or tree head.

You can run a single classifier for debugging or a long zoo list for a bake-off. The report writer later flattens those per-method results into a table.

Optional classifier optimization

When optimization.enabled is true and the method is a neural network that supports the hooks, ARCADE can run pruning, distillation, neural architecture search, quantization, or Optuna search after or around the primary training path. Optimization is intentionally optional. Many classical methods have nothing to prune. Leaving the section disabled keeps first runs short and keeps the report focused on accuracy and hardware estimates alone.

Hardware estimation

The hardware stage turns model structure into analytical estimates of look-up tables, DSP blocks, and latency under stated word-length assumptions. These numbers are meant for early comparison across methods, not for claiming board-level timing closure. Treat them as order-of-magnitude guidance that helps you discard methods that are clearly too large or too slow for your target fabric.

Optional quantum error correction curves

When the QEC section is enabled, ARCADE maps readout error and readout duration into logical error rate versus distance and into wall-clock cycle time for memory-style experiments. This stage answers a systems question: does a slightly worse discriminator that finishes much faster improve the fault-tolerant budget? The curves are only as meaningful as the noise model and decoder assumptions you declare.

Visualization and report generation

The final stage writes figures and a summary report into the configured output directory. The report is the artifact you should read first after a zoo run. It collects joint accuracy, per-qubit accuracy, F5Q where applicable, latency, and FPGA estimates into one place so you do not have to dig through every method subdirectory to answer which method won under the shared recipe.

Mapping stages to packages

Stage	Primary package area	Config section
Data preparation	`arcade.data`, `arcade.features`	dataset, demodulation, split, features
Classification	`arcade.classifiers`, `arcade.stages.classify`	classifiers
Optimization	`arcade.classifieroptimization`	optimization
Hardware	`arcade.hardware`	hardware
QEC and FTQC	`arcade.qec`	qec
Report and figures	`arcade.visualization`	visualization

Exact field names and nested schema live in Config schema. Registry names live in Classifiers reference.

Next

Use Verification to confirm your install before a long run. When you are ready for a guided bake-off, open First zoo run.