Shared recipe

Calling a method the best classifier is meaningless without a frozen bake-off recipe. ARCADE’s headline claim is fairness: the same traces, the same preprocessing, the same splits, the same metrics, and the same report shape for every method in a given run.

What is frozen

Ingredient	Where it lives	Why it matters
Trace file and keys	`data.path`, `data.hdf5_keys`	Everyone sees the same shots
Demodulation, boxcar, truncate	`data.demodulation`, `data.preprocessing`	Changes the IQ geometry
Label packing	`data.label_format`, `data.qubit_bit_order`	Bit-order bugs look like bad models
Train, validation, and test	`data.split`	Leakage across methods is silent
Filter bank	`classifiers.filter_types`	MF, RMF, and EMF must be shared unless a method opts out
Metrics and report	`visualization` settings	Comparable tables

The public reference eight-method recipe is configs/paper_all_methods_benchmark.yaml. Dedicated configs/paper_*.yaml files may differ slightly when a paper needs a different protocol. Those are paper recipes, not silent edits to the reference benchmark.

What is allowed to vary

Which names appear in classifiers.run. Per-method option blocks such as fnn.config_file. Optional stages such as optimization, qec, and hardware.run_synthesis.

Pitfalls checklist

HERQULES zoo filters are typically [MF, RMF]. Claiming bit-exact paper parity needs the paper YAML. Path-signature zoo configs may set time_augmentation: true while the paper path is two-dimensional I and Q. Treat Readout 2019 path-signature results as transfer under the zoo recipe. The name mlp is the scikit-learn multilayer perceptron. It is not the Lienhard-style fnn. The name reservoir is a deprecated alias of ngrc. It is not a classical echo-state network.

Changing the recipe between methods turns the leaderboard into marketing. Keep knobs under data once. Swap only classifiers.run when you want a fair comparison.