Editorial policy — no invented numbers

The rules

Every catalog figure carries a source, a date, and a method (recorded as VENDOR_UNSPECIFIED when the vendor did not state one). No exceptions, not even temporarily.
Lessons may use hedged order-of-magnitude characterisations (“around 1%”) only when linked to the sourced catalog for per-device figures; specific benchmark results always carry their own source.
We never fabricate, estimate, interpolate, or borrow a missing specification from a sibling device.
“Not publicly disclosed” is a finding, recorded with the date we checked and where we looked — never a blank cell.
Vendor claims are labeled as vendor claims. Most are accurate; a claim and an independent measurement are still different kinds of evidence.
Methods are part of the number. A median randomized-benchmarking fidelity is not comparable to a best-pair figure, and our comparison engine says so rather than hiding it.
Theory, lab demonstration, hardware demonstration, practical capability, roadmap claims, and disputes are never blended — in a sentence, a table row, or a headline.
No numeric hype scores. Classifications with dates, sources, scales, and hardware — or nothing.
No universal “best QPU” ranking, ever. Contextual verdicts only.
No language model writes a hardware specification. Every figure traces to a human-checkable public source.
Corrections are public. Every reported error is logged at /corrections, whether or not it turns out to be right.

You can click any figure on this site and see where it came from, when, and how it was measured. If you cannot, that is a bug — report it.