# Metrics

The comparison table in the summary report is only useful if you know what
each column measures. The definitions below match what ``arcade.viz.report``
writes into ``summary_report``. Read this page before you treat a ranking as
final, especially on multi-qubit joint labels where a single wrong bit order
can quietly destroy the headline number.

## Joint accuracy

Joint accuracy is the fraction of shots where the full multi-qubit label
matches the ground truth. For five qubits with joint labels this is a
thirty-two class problem. Joint accuracy is the primary ranking key in the
recommendation box on the executive report page.

A wrong ``data.qubit_bit_order`` systematically destroys joint accuracy while
per-qubit numbers can still look plausible. The zoo default is ``lsb0``.
Several paper YAML files use ``msb0``. Do not mix those conventions silently
across configs you intend to compare.

## Infidelity

Infidelity is typically one minus accuracy, or a closely related report
variant. It is useful when you think in error rates, for example when binding
a classifier score into stim as a readout error probability.

## F5Q

F5Q is an aggregate of per-qubit fidelities, using a product or
geometric-style combination in the zoo report. F5Q can look healthy even when
one qubit is weak if the others are excellent. Always read it next to PQ min
before declaring balance across the device.

## PQ min and per-qubit fidelity

PQ min is the worst single-qubit assignment fidelity. Per-qubit fidelity
columns show each qubit separately. Use these to spot imbalance. A method can
win joint accuracy while harming one qubit that your algorithm cares about.

## Hardware columns: LUT, DSP, and latency

These columns are analytical estimates from ``arcade.hardware``. They are not
post-place-and-route guarantees. They exist so you can reject a half-percent
accuracy win that costs an order of magnitude more look-up tables. Scope and
limits are explained in {doc}`hardware_limits`.

## Duration and sweep curves

When ``sweep.enabled`` is true, ARCADE truncates traces at several lengths and
plots speed against fidelity. Shorter readout usually raises error. The curve
is how you find a knee before binding into QEC cycle time. See
{doc}`ftqc_bridge` for that next layer of reasoning.