ARCADE Readout Zoo
Home Learn Classifiers Results Quickstart Docs FTQC and NISQ goals

ARCADE

Readout Zoo

A shorter measurement that loses a little assignment accuracy can still improve logical error rate

Published readout methods rarely share an experimental protocol. ARCADE provides a common evaluation recipe, places each method in one architectural pipeline, and connects classifier performance to the system goal you care about — fault-tolerant cycle budgets or a NISQ operating point for later pulse design.

Start with readout errors See FTQC and NISQ goals

Purpose

Why a common evaluation framework

Every published discriminator can look strong on its own traces. Deployment requires a comparable ranking under one protocol, resource estimates, and an answer to what that ranking means for NISQ circuits or fault-tolerant QEC cycles.

Every method shares the same prepared traces. Features are built for the method’s kind — filter scores, raw IQ, or a hybrid expansion — then a single classify block runs.

Optional compression and tuning stay behind switches. Fabric and latency estimates, then FTQC or NISQ system checks, consume the same prediction stream. You change the method block — not the experiment.

Classical thresholds, Lienhard-style networks, HERQULES-style hybrids, and next-generation reservoir features are therefore comparable under one protocol.

Prepareshared traces
Featuresfilters · IQ · hybrid
Classifymethod block
Optionalcompress · tune
Estimatefabric · latency
System checkFTQC · NISQ

Primary focus for this project

Fault-tolerant quantum computing

Map assignment quality to readout error. Scale idle noise with acquisition time plus post-measurement control overhead. Report logical error rate versus duration and wall-clock cycle time = readout + control + decode.

Logical error and cycle time →

Also in scope

NISQ operating points

From the classifier side: accuracy versus readout latency. Choose an operating point under your coherence budget; that point can later guide readout-pulse optimization.

NISQ framing →

Reference evidence

One protocol. Many published methods.

The public reference comparison uses a frequency-multiplexed five-transmon superconducting readout dataset and ranks classical, hybrid, and neural methods under identical splits and metrics.

How to read this plane

Each point is one method under the shared five-transmon protocol. Higher is better assignment accuracy; left is cheaper analytical fabric cost (LUT).

Classical filter cuts sit near the low-cost side. Learned and hybrid methods climb accuracy at higher estimated fabric cost. Accuracy alone is not a deployment criterion — latency and the FTQC or NISQ ending still matter.

Open the full comparison →

Accuracy vs fabric cost
Accuracy versus estimated fabric cost

Marker size scales with estimated latency.

Shared evaluation recipe

What a fair comparison freezes

Methods disagree on architecture. They should not disagree on the experimental protocol when the goal is ranking. The shared recipe locks preparation, splits, and metrics so only the classifier block changes.

  1. Qubit system. Public reference: frequency-multiplexed five-transmon superconducting readout. Paper recipes may use their own traces — labeled as such.
  2. Preparation. Demodulation and preprocessing are shared so every method sees comparable IQ evidence.
  3. Splits. Train and validation partitions are frozen for the reference ranking.
  4. Features. Built according to each method’s kind — filter bank, IQ, or hybrid map — from the same prepared traces.
  5. Metrics. Joint and per-qubit assignment quality, plus analytical fabric and latency estimates when enabled.
  6. System endings. Optional FTQC logical-error curves and NISQ accuracy–latency views from the same predictions.