Methods

Every number published by Cittela carries five things inline, at the point where the number appears:

the model it was measured on, the arm it is compared against, the gate it was required to pass, the scope within which it holds, and its unit.

A number missing any of these is a draft, not a result, and does not appear on this site or in the deposits.

Pre-registration here means the acceptance gate for a measurement is written down before the measurement is run: what will be measured, on what arms, and what threshold counts as passing.

A figure that fails its gate is reported as a failed gate. It is not re-scoped after the fact into a weaker claim that passes.

When a figure published by Cittela is superseded or withdrawn, it is recorded here by name, with the reason and its replacement, so that no earlier copy of a claim outlives its correction.

Entries, August 2026:

Flat rotation-cost offset. The claim that per-tenant rotation costs a fixed number of seconds per request was withdrawn in August 2026 as unmeasured, not refuted: no run had isolated it. Superseded by the v2.4 result that the rotation cost is prefill work, scaling approximately proportionally with prefix length. It is reported as a fraction of prefill, never as flat seconds.

matched_pool protocol v2.1 and v2.2. Superseded by later protocol revisions; v2.4 is current. Figures measured under superseded protocol versions do not carry forward without re-measurement under the current protocol.

Other performance figures that circulated in earlier external material were withdrawn in August 2026 pending re-measurement and were never published on this site. The density figures on the home page are v2.4 figures and are current.