research build log · ARC-AGI-3 · weekly snapshots

▣ code‑native/arc‑agi‑3 — archive

An agent that plays ARC‑AGI‑3 games by writing Python against the frame, with an execution-grounded “sleep” phase that mines skills and lets the game itself be the judge. Each card below is one week’s snapshot of the log — newest first.

experiment journal (2026-07-09): ▸ 무누수 파이프라인 플로우차트  ·  ▸ RQ 카드 저널

2026‑07‑11 latest

TTSO: 무누수 첫 풀게임 WIN 더블 — ft09 6/6 ×2, fork=0

엔진 복제(fork) 없이 predict-verify만으로 ft09 전 레벨을 두 독립 런이 클리어. 인터랙티브 파이프라인(노드 클릭 = 실제 프롬프트/트레이스), 여섯 위기 타임라인, 선행 4-레인. (07-09 통합판은 이 페이지로 흡수 — 원문은 git 이력에 보존.)

6/6 levels × 2 runs fork_dependence = 0 RHAE 40.9 / 55.7
read the week →
2026‑07‑03 fork-시대 스냅샷

v2: sleep harness, escapes 0 → 4 levels, RHAE, trace explorer

A harness coordinate bug was silently eating every click. Eight fixes later, the agent clears level 4 on ft09 — with a data-driven run viewer over the real traces. 정정(07-09): 이 “WIN”은 이후 fork-브루트포스 아티팩트로 판정 — 역사 기록으로 보존.

level 4 cleared (fork-era) RHAE 17.9 21 runs · 8 fixes
read the week →
2026‑06‑30

v1: code-native baseline, failure ladder, first architecture

The first honest log: one autonomous CodeAct loop, an LLM-free skill gate, and the failure-mode chain — built and genuinely code-native, not yet solving a level from scratch.

0 levels from scratch failure-mode ladder LLM-free skill gate
read the week →