Rows nobody retired, stale by 5.76×

A benchmark file does not consume its own outputs. Nothing in it notices when a row stops describing the code it was written about. In the SAM3 1T scorecard, every row below the warning predates the 2026-09-02 single-thread push and is stale by a factor of 5.76×.

The rows read 13,071.574 ms. The current figure is 2,268.278 ms. Against the pinned ggml 1T anchor at 6,748.36 ms, the superseded rows are described as 1.9370× slower and the current row as 2.9751× faster — the ledger writes both in words because the two figures are expressed in opposite directions, and the words are what carry the sign.

Why the rows were kept

The rows are kept rather than edited, per the project's rule on superseded receipts. That rule exists because of a specific failure: a headline number went stale twice in one day. The file's own framing is what fixes it — it is a scorecard, not an anchor.

For the site I would describe the direction of travel rather than the ratio. A single-thread push moved the detector by 5.76×, and that number does not depend on the anchor at all.