Stale "best known" targets in AI-math benchmarks

How many stored targets in public benchmarks of AI mathematical discovery have been superseded by the literature, or cannot be scored as written. Every figure in the tables is computed from the per-target audit files released with the paper (CC BY 4.0).

…

Per benchmark

BenchmarkPinnednStale or unscoreableWilson 95%IntervalUNCLEARStale at pinMedian lag to pin (days)Last audited

"Stale or unscoreable" = STALE + MALFORMED. UNCLEAR and leaderboard-only rows are shown, never counted. "Stale at pin" and "median lag" use a mechanical rule on the released files: a STALE row whose source date (the first full date given) precedes the pin date; the lag is the pin date minus the source date. The paper's hand-checked stale-at-pin counts, which also read dates given in the notes, are 5 (notebook) and 8 (paper; one of them now UNCLEAR). HorizonMath's pinned commit has no stated date, so both cells are blank for it.

Non-current rows

Non-novel announcement rate per claimant

Registered forecast

Registered 2026-10-12, before any of the objects exist. Window: releases between 2026-10-13 and 2027-04-12. Objects: the next HorizonMath release or data update (at least 10 changed rows); the next AlphaEvolve results release (at least 10 new rows, or a paper); the next EinsteinArena results paper or announced snapshot; any new public AI-math benchmark with at least 30 numeric "best known" targets.

Adjudicated per object 21 days after its release, hard stop 2027-05-03. Thresholds do not move. A negative result is reported as "the field corrected itself"; no qualifying release by 2027-04-12 is reported as "no observation". Status: open.