Why a result says what it says
This platform prints an explanation beside most of its numbers. Each one follows a rule, and the rules are worth more than the instances — every one of them exists because the natural reading of a number is wrong in a specific way. Each links to a page where you can see it in force.
Every number says whether it was measured, computed, or assumed.
Without it, a reader concludes: Reading a modelled figure as an observation. A noise assumption and a sampled result look identical once both are printed to three decimal places.
The mitigation panel carries a basis column on every row, so a modelled correction is never mistaken for a measured one.
Not observing something is not the same as it not happening.
Without it, a reader concludes: Reading zero observed failures as a rate of zero — and a zero-failure row sorts to the top of every ranking, so it is the row most likely to be quoted.
A run with no failures reports a Wilson upper bound instead of a rate, because 0 in 10,000 shots bounds the rate near 3.8×10⁻⁴ and says nothing below that.
Two numbers are not a comparison unless the same things were held fixed.
Without it, a reader concludes: Treating a single merged leaderboard as authoritative. It is the artifact that makes an invalid comparison look settled.
Runs measured at different distances, error rates, noise models or suite versions are grouped separately and never ranked together.
A check that could not decide must say so, not fall back to a verdict.
Without it, a reader concludes: Reading the absence of a counterexample as a proof. Most verification returns 'not shown', which is not the same as 'shown false'.
ZX equivalence reports a strict INCONCLUSIVE distinct from both proved and refuted, and produces a counterexample when it refutes.
Getting the right outcome is not the same as getting the expected answer.
Without it, a reader concludes: Scoring a decoder by whether it named the error. On a degenerate code it routinely names a different one and restores the state perfectly.
A correction succeeds when the residual is a stabilizer, so six of Shor's twenty-seven single-qubit errors are corrected under another name.
When a result depends on a chosen constant, the choice is part of the result.
Without it, a reader concludes: Quoting a resource estimate as though it were determined by physics. Two equally conventional prefactors change the machine size by 2.78×.
The fault-tolerant panel shows both prefactors with the one in use marked, and states the difference as a ratio rather than a distance.
What was measured outranks what was declared.
Without it, a reader concludes: Treating metadata as evidence. An author typing a code family into a form is not the same as runs demonstrating it.
Recommendations weight measured evidence above every metadata claim, and exclude popularity entirely — stars measure attention, not correctness.
Self-consistency is not the same as provenance.
Without it, a reader concludes: Reading a matching reproducibility hash as proof the numbers came from where the record claims. Anyone can hash a fabricated result correctly.
Execution class is a separate claim from the hash, and results labelled HARDWARE are refused because this registry cannot verify them.
New to these terms? The glossary defines each with the mistake it prevents, and the guided tour walks several of these principles as things you do rather than read.