Recursive Self-Improvement for ARC-AGI-3: Hill-Climbing to 78.95 RHAE with Haiku 5.5 and ASSAY

A small model can be made to reason better on a hard interactive benchmark without touching its weights, by improving the harness it reasons in. We report a human-in-the-loop recursive self-improvement loop around ASSAY, a general reasoning harness for coding agents whose layers of reasoning abstraction locate where reasoning breaks. The loop runs the harness on the 25 public games of ARC-AGI-3 and hands the complete record to a strong analyst model, Claude Fable 5.1, which reads every loss through the layers and recommends code changes at the layer each loss points to. The author adopts every recommendation on a tagged branch, and the loop runs again. Because ARC-AGI-3 verifies a win, a won game is pinned to the version that won it and the newest version attacks the rest. After two rounds, the newest version on its own won 17 of 25 games with Claude Haiku 5.5, at a set RHAE of 76.26, with one medium-effort pass over all games and one xhigh pass over the 13 it lost. Those 17 are every game any earlier version had won. The pinned record over six passes, which keeps each game's best verified run, is 17 of 25 at 78.95. At medium effort alone, Haiku went from 5 of 25 and 37.18 under plain Claude Code to 12 of 25 and 60.24 under the twice-fixed harness (paired +20.4 over the unmodified harness, seven wins gained and none lost, p = 0.016). Haiku's model fit every transition it tested and was wrong in regimes it never tested; its predictions there were too vague for the harness's meters to see the error; then it stopped. The first round moved the stop one session later, and the second removed the exits the first had opened. All scores are computed locally over several attempts on public games, and are not comparable to ARC Prize's single-attempt verified results.