NeuroGolf 2026 — Root

Minimize ONNX neural networks for 400 ARC-AGI tasks. Score = sum of per-task scores (max 10,000).

Folders

Folder	Purpose
data/	Raw task JSONs (task001.json … task400.json) — downloaded from Kaggle, not in git
knowledge/	Per-task knowledge base (task_001/ … task_400/) — persistent memory for agents
submissions/	Submission zips and ONNX bundles, versioned by score (s_6256_03/, staged/)
tools/	CLI scripts for scoring, staging, submitting, and inspecting tasks
dashboard/	Flask web dashboard — 400-task color grid + task detail pages

Key files

File	Purpose
AGENTS.md	Read first. Codex agent SOP — env, scoring formula, 9-step workflow, pitfalls
strategy.md	ONNX pattern library (A–J), memory tricks, per-task score history
local_score.py	Reproduces official Kaggle scoring locally (6256.04 ≈ 6256.03 LB)
requirements.txt	Exact pinned library versions matching Kaggle scoring environment
competition_description.md	Competition rules, scoring math, critical constraints

Quick start

conda activate neurogolf

# Dashboard
python dashboard/app.py       # → http://127.0.0.1:5001

# Find worst tasks
python tools/list_tasks.py --bottom 30

# Inspect a task
python tools/show_task.py 233

# Score a candidate ONNX
python tools/score_task.py 233 candidate_233.onnx

# Stage + submit
python tools/stage_submission.py 233 candidate_233.onnx
python tools/submit_task.py 233

# After LB confirms improvement
python tools/confirm_lb.py 233 <new_total>
git push origin main

Environment

• Python: /Users/yeyang/miniconda3/envs/neurogolf/bin/python
• Root: /Users/yeyang/Desktop/uzh26s/golf/
• Scoring: score = max(1.0, 25.0 - log(max(1.0, params + memory_bytes)))
• Baseline: 6326.10 (LB, 2026-06-07)