EOSE Labs · ARC-AGI Floor

We're building toward 64%.
Come help.

ARC-AGI tests whether AI can actually reason — not memorize. We're building a 3-cap execution verifier on sovereign infrastructure. The floor is open. The standards are public. We want builders.

0%
64% target
〰️ curious welcome 🌀 wonder encouraged ⚓ no sorry needed
Ways to contribute
⚙️
Code · Fix a gap
Run the executor. Find where it fails. Fix it. The parse loop, the vote logic, the color equivariance layer — all open problems with clear success criteria.
Python Ollama ARC tasks
→ arc-3cap repo
◈
Cap builder · Add a stance
Design a 4th prompt identity. Each cap is a different way of looking at the same puzzle. More stances = better vote diversity = higher score. Test it. File it as a standard.
NP standard floor
→ floor@pemos.ca
📐
Standard builder · File a pattern
We file everything we learn as a named pattern standard (NP). If you crack something — parse reliability, voting logic, color equivariance — it becomes a filed standard, not just a code comment.
NP ARB LSOS
→ see the floor
🔬
Tester · Run and report
Download the runner. Run it on your hardware. Report what you see — which task types fail, which caps do better, where the vote breaks down. Raw results are as valuable as fixes.
results validation
→ audit@pemos.ca
🌐
Community · Bring others
Post results. Share the floor. Point people at the viz. The ARC Prize community is 500+ people who care about real reasoning. We're building something worth showing.
ARC Discord LinkedIn
→ pemos.ca/arc-agi
⚖️
Advisor · Shape the floor
If you've worked ARC-AGI seriously, we want to hear how you think about task selection, failure taxonomy, and what beats 64%. CLO-gated. Serious only.
C=11+ CLO gate
→ legal@pemos.ca
Open gaps · pick one up
GAP-PARSE
Reliable grid extraction
P0
LLM output → valid [[int]] grid, reliably, every time. Current: ~60% parse rate.
GAP-COLOR
Color equivariance
P1
Permuting colors shouldn't change the answer. Each cap tries a different color mapping. Remap back at end.
GAP-VOTE
Vote with confidence
P1
3-cap vote · when no majority, pick the cap with best training accuracy. Don't withdraw blindly.
GAP-TASK
Task selection
P2
Some tasks are solvable by small models. Predict which ones. Spend budget on winnable tasks.
GAP-DSL
DSL primitives
P2
ARC patterns recur. A small library of grid primitives (mirror, rotate, fill, mask) helps caps reason faster.
Where we sit · ARC-AGI v4
Team Score · Note
1
noemon.ai
91.4%
LLM+DSL+vote
2
EdisCore
84%
transform() verifier
3
P'tit Ju
57.6%
DSL+TDD
→
EOSE Labs (3-cap verifier)
→ 64%
building now
—
compress-arc
20%
floor to beat
How we work · the 3 says
〰️
Wiggles
Curiosity is enough to start. You don't need to know everything. Come look. Ask questions. That's a contribution.
🌀
Wonder
The best ideas come from people who look at a problem sideways. Don't suppress the odd angle — that's often where the gap closes.
⚓
No sorry
No gatekeeping. No hierarchy of worthiness. If your contribution is real, it goes on the floor. The work speaks.

The floor is open.

We're building toward 64% on ARC-AGI v4. Come check the floor, pick up a gap, or just watch what happens. All paths welcome.

→ see the floor floor@pemos.ca lookbook