The measured succession floor: why a new release was presumed at least as good as the one it replaced

Until 4 September 2026 the ModelCap board ranked on the lower end of an uncertainty interval. That suited a board that wanted to be conservative, and it had one predictable side effect: a fresh release with a handful of boards and no Arena rating had a wide interval, so its lower bound could land beneath an older, thoroughly measured release it beat on every board they shared. For a day or a week the table then said the old model was better, which nobody who read the evidence believed.

Index version 4.8, deployed on the evening of 1 September 2026, answered that with a presumption: a measured successor was at least its measured predecessor's equal until like-for-like evidence said otherwise.

The rule

• The row had to be measured, current, canonical, and in a product family with an older sibling that was also measured. Rows with no evidence used the separate succession start rank from version 4.7; rows placed from a model card were not eligible.
• If the successor's lower bound sat beneath the predecessor's score, it was floored at the predecessor's score plus 0.1. The point estimate was never changed and could sit below the floor. The top of the interval became the higher of its own ceiling and the predecessor's, so a thin successor inherited a ceiling instead of collapsing onto a zero-width line.
• The floor was withdrawn when, over the scored boards that had measured both models, the predecessor's weighted share exceeded the successor's. Each shared board counted with its family weight times its source share, using the best score per board. No shared board at all meant the presumption stood.
• The row was labelled Measured · succession and published the predecessor id, the floor, the shared boards and both shares, so the presumption was auditable from the row itself.

Who counted as a successor

Order came from the catalogue's declared supersession edge first, then from a higher generation token in the name, then from a release at least thirty days later. The one trap we hit building this: dated refreshes of the same product, such as DeepSeek V4 Pro 0423 and 0813, tie on the generation token. Only the declared edge ordered them, and the family resolver that writes that edge ran before the index, so it was always present when the index needed it. Chains of three or more generations were settled by iterating to a fixed point, so the order in which the catalogue happened to list the rows could not change the result.

What it did to the board

We previewed the change against the live corpus before shipping it by replaying the sealed version 4.7 index (zero drift against production, which is the precondition for trusting the preview) and diffing it against 4.8. Seven rows were floored.

The DeepSeek rows are the striking ones. Each was a dated refresh that had retired a seven-board predecessor and was itself measured on a single ARC-AGI-2 row, so its interval was enormous and its lower bound sat in the twenties while the model everyone was actually using sat in the archive. The Fable row is the small one, and it is the one people noticed; the arithmetic is in the launch note.

Keeping it honest

Two guards kept the floor from becoming a way to launder a rank. The data validator normally insisted that a point estimate sat at or above its lower bound; for floored rows it instead cross-checked the predecessor row, so a floor with no measured predecessor behind it failed the build. And the scoring code that decided whether shared boards favoured the predecessor used the same family-weight and source-share plan the score itself used, so the withdrawal test could not drift from the score. The rules in force today are in the methodology.