Read the report

The zoo run is finished. Somewhere under visualization.output_dir sits a summary report and a results.json. This tutorial treats that report like a case file: open it, read the headline table, check the traps, then decide what to ship or what to rerun.

Mission. Open summary_report.pdf or HTML, explain joint accuracy versus F5Q versus PQ min in one sentence each, and say whether the recommendation box survives the hardware columns.

        flowchart TB
  OUT[output directory] --> PDF[summary_report]
  OUT --> JSON[results.json]
  PDF --> TAB[Comparison table]
  TAB --> REC[Recommendation box]
  TAB --> HW[LUT · DSP · latency]
  REC --> DEC{Ship, tune, or rerun?}
  HW --> DEC

Where to look

After arcade run finishes, open the directory named by visualization.output_dir. A typical layout looks like this:

{visualization.output_dir}/
├── summary_report.pdf
├── results.json
├── data/
├── filters/
├── {classifier}/
├── sweep/
└── hardware/

data/IQ scatter, averaged traces
filters/MF / RMF envelopes
{method}/confusion, training curves
sweep/speed vs fidelity
hardware/resource comparison

Configuration knobs that change report shape include visualization.summary_sections, visualization.report_detail, and visualization.format. Supported formats include pdf and html.

Anatomy of report detail tiers

Tier	`report_detail`	Contents
Executive	`summary` and above	Comparison table with Accuracy, Infidelity, F5Q, PQ min, LUT, DSP, and Latency in nanoseconds; confusion small-multiples; speed and fidelity overlay; recommendation for highest joint accuracy
Detail	`standard` and above	Per-qubit fidelity table; per-classifier confusion and hardware blurbs
Appendix	`full`	Cross-fidelity heatmap; grouped per-qubit bars

How to read the comparison table

Read the table left to right as a story, not as a single scoreboard number.

Accuracy joint is the fraction of shots where the full multi-qubit label matches. On five qubits this is a thirty-two class problem. It is harder than averaging per-qubit scores. This is usually the ranking key in the recommendation box.

F5Q is a geometric or product-style aggregation of per-qubit fidelities. It is useful when joint accuracy is dominated by one weak qubit. Read it next to PQ min.

PQ min is the worst single-qubit fidelity. It flags imbalance across the device. A method can win joint accuracy while quietly harming one qubit your algorithm cares about.

LUT, DSP, and Latency are analytical FPGA estimates from arcade.hardware. Treat them as ranking aids, not bitstream guarantees.

Recommendation box highlights the highest joint accuracy under this shared recipe. Always check the hardware columns before deciding what to ship.

Quick read order. Joint accuracy → PQ min → F5Q → hardware columns → recommendation box. If any step disagrees with the box, write down why before you trust the ranking.

Pitfalls checks

Warning

A wrong data.qubit_bit_order systematically destroys joint accuracy while per-qubit numbers can still look plausible. The zoo default is lsb0. Path signature results on Readout 2019 under the zoo recipe are transfer results. Analytical LUT wins are not Vivado timing closure. See Shared recipe and Metrics before claiming paper parity.

Next

Read Metrics for formal column definitions, Hardware limits for FPGA caveats, and Classifiers reference for how each method earned its row. When you are ready to add your own row, open Build your own classifier.