GoKawiilDrop7 AI project reports learned evaluator beating fair-leaf search in screen-tier test, replicated twice
A hobbyist research log on the Drop7 puzzle game describes a learned evaluator — lookup tables trained via temporal-difference self-play for four billion moves — outperforming a 'fair leaf' baseline inside the same depth-3 search. In a 256-game test it scored 484,577 points to 314,438, and the result replicated on a separate 512-game block on 2026-09-06 with a verified identical model, scoring 487,066 to 326,717.
GoKawiil's interpretation of the reporting above, not reported fact.
The author is explicit that this is a preliminary 'screen tier' result on public development seeds, not a verified qualification, and that the mean score is still under half the project's target — suggesting the milestone is promising but far from conclusive. The careful separation of training-selected validation scores from blind screen scores, and the disclosure that a diagnostic comparison was not the preregistered test, points to an unusually rigorous, self-skeptical approach to evaluating game-playing AI claims.
- A learned lookup-table evaluator beat a fair-leaf baseline search in Drop7 across two independent test blocks.
- Results remain at 'screen tier' on public development seeds, not a final or protected benchmark.
- The author distinguishes honest blind-test margins from inflated validation-selected margins to avoid overstating progress.
Source: drop7.dev, 2026-10-23
Published there as: “Show HN: Searching for optimal strategies in a puzzle game with chance (Drop7)”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.