[schema]: Frontier Models with the Right Harness Achieve ~99% on ARC-AGI-3 Public

Summary

Schema harness (not weight changes) for ARC-AGI-3: observe→deliberate→execute→record with an append-only timeline. Deliberation writes step(state,action), run_backtest over all recorded transitions, run_bfs inside the certified program, commit_actions only. Self-reported Public-set RHAE 98.98% with Opus 4.8+Fable 5 fallback (vs 42.83% same pairing in Claude Code) and 95.35% with GPT-5.6 Sol xhigh/max fallback. Official Sol max was 13.33% Public / 7.78% Semi-private — not a matched harness ablation. Scores unverified by ARC Prize. 50 released trajectories (25+25). Residual: Claude 19/25 games at 100 RHAE, six remaining 89.87–99.10; Sol 20/25 at 100.

Keywords

schema · arc-agi-3 · harness · world-model · rhae · impossible-research · blog · x

Topics

ARC-AGI, agent harnesses, symbolic world models

Research notes

• Primary: Impossible Research blog (cite schema2026). Discord/X https://x.com/HavenFeng/status/2077770348876247502 via fxtwitter. Traces https://huggingface.co/datasets/schema-harness/arc-agi-3-schema-traces → datasets_local ARC-AGI-3 Schema Gameplay Trajectories. Earlier VIGA (state grounding) / WorldCoder (transitions) cited as one-sided predecessors. Self-reported public-set only.