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

The ModelCap board ranks on the lower end of an uncertainty interval. That is the right statistic for a board that wants to be conservative, and it has one predictable side effect: a fresh release with a handful of boards and no Arena rating has a wide interval, so its lower bound can land beneath an older, thoroughly measured release it beats on every board they share. For a day or a week the table then says the old model is better, which nobody who reads the evidence believes.

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

The rule

• The row must be measured, current, canonical, and in a product family with an older sibling that is also measured. Rows with no evidence use the separate succession start rank from version 4.7; rows placed from a model card are not eligible.
• If the successor's lower bound sits beneath the predecessor's score, it is floored at the predecessor's score plus 0.1. The point estimate is never changed and may sit below the floor. The top of the interval becomes the higher of its own ceiling and the predecessor's, so a thin successor inherits a ceiling instead of collapsing onto a zero-width line.
• The floor is withdrawn when, over the scored boards that have measured both models, the predecessor's weighted share exceeds the successor's. Each shared board counts with its family weight times its source share, using the best score per board. No shared board at all means the presumption stands.
• The row is labelled Measured · succession and publishes the predecessor id, the floor, the shared boards and both shares, so the presumption is auditable from the row itself.

Who counts as a successor

Order comes 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 orders them, and the family resolver that writes that edge runs before the index, so it is always present when the index needs it. Chains of three or more generations are settled by iterating to a fixed point, so the order in which the catalogue happens to list the rows cannot 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 keep the floor from becoming a way to launder a rank. The data validator normally insists that a point estimate sits at or above its lower bound; for floored rows it instead cross-checks the predecessor row, so a floor with no measured predecessor behind it fails the build. And the scoring code that decides whether shared boards favour the predecessor uses the same family-weight and source-share plan the score itself uses, so the withdrawal test cannot drift from the score. The full contract is in the methodology.