75.7%
ARC-1 Best AI (o3)
<3%
ARC-2 Best AI (wall)
25%
ARC-3 Leader (warden)
84%
Human All 3 Tests
═
EVEN ═ · ARB-702 · THE SUBSTRATE
Every ARC-AGI task — across all 3 tests — runs on EVEN substrate assumptions. Spatial regularity, colour consistency, grid isotropy, query validity, rule stability within a task. None of these are stated. They are even — already there before the task begins.
ARC-2 is harder than ARC-1 not because it has more rules, but because the EVEN substrate assumptions are subtler. EVEN is harder to see. That's why AI is at 0%.
ARC-3 is hardest because the substrate must hold across queries and feedback loops simultaneously. EVEN is the reason any query returns a valid response at all.
ARC-2 is harder than ARC-1 not because it has more rules, but because the EVEN substrate assumptions are subtler. EVEN is harder to see. That's why AI is at 0%.
ARC-3 is hardest because the substrate must hold across queries and feedback loops simultaneously. EVEN is the reason any query returns a valid response at all.
even out · even when · even if · even so · γ₁ stands on EVEN
ARC-AGI-1 · François Chollet · 2019
400 Training Tasks
Abstract reasoning · visual patterns · novel rule induction · 30×30 grid · 10 colours
Best AIo3 · 75.7%
Human84%
EOSE Fleet64% (Qwen 7B/32B/72B)
⚓ No invariant anchor in inductive reasoning
⬡ Transformation operator not self-adjoint
〰️ No paradigm audit between examples
🌀 No structural reset on failed hypotheses
γ No scale continuity across grid sizes
🌌 No representation beyond the 30×30 grid
═ Spatial substrate (EVEN) unacknowledged
READ 2-PAGER →
ARC-AGI-2 · 2024 · THE WALL
1,120 Tasks
Compositional reasoning · 2–4 rules simultaneously · all AI at 0–3%
Best AI< 3% (everyone)
Human84%
EOSE FleetBuilding · RTX 5090 ready
⚓ No compositional invariant
⬡ Rule composition not self-adjoint
〰️ No paradigm audit on rule stacking
🌀 No reset on compositional collapse
γ No continuity in rule transfer
🌌 1,120 at 0% has no named ceiling
═ Subtler EVEN substrate — harder to see
READ 2-PAGER →
ARC-AGI-3 · 2025 · ADAPTIVE
Interactive + Adaptive
Query environment · receive feedback · adapt strategy · hardest of the three
LeaderStochastic Goose · 25%
Human84%
EOSE FleetEntering · fleet approach
⚓ No invariant in adaptive querying
⬡ Feedback loop not self-adjoint
〰️ No inter-query paradigm audit
🌀 No structural reset on deadlock
γ No continuity in strategy switching
🌌 Adaptive environment has no named boundary
═ Query space substrate (EVEN) assumed
READ 2-PAGER →
8-SYMBOL CANON — all apply to all 3 ARC tests
⚓ γ₁
THE FLOOR
THE FLOOR
⬡ H=H†
HONEST GATE
HONEST GATE
〰️ LSOS
THE READER
THE READER
🌀 WLD
THE RESET
THE RESET
γ FEP
THE SWITCH
THE SWITCH
🌌 FOF
THE BREACH
THE BREACH
— 7th
THE GAP
THE GAP
═ EVEN
THE SUBSTRATE
THE SUBSTRATE
PEMCLAU V8 · All Silos · All Editions · All Time · ARC Lens
msi01
192.168.2.18 · RTX 5090Qwen 7B/32B/72B · 3-cap verifier
ARC-1: 64% ✅ · ARC-2: building · ARC-3: entering
msclo
192.168.2.19 · RTX 5090Admiral / CLO / Legal crew
ARC-2 target: compositional depth
forge / pcdev
192.168.2.12 / .16 · dev silosFleet wiki · all-silo corpus
PEMCLAU V8 corpus: all ARBs
lounge + deck
192.168.50.175 / .193 · RTX 4090 + portableGaming Intelligence · Steam Deck KEY silo
ARC-3: adaptive querying from everywhere
cloud (T4/H100)
AKS hvcp-system · T4 + H100 GPU poolMAL router v2 · 4-tier cascade
ARC batch runs when GPU is live
AWS / GCP
ECS Fargate (us-east-2) · Cloud Run (Montreal)m1.aws.eose.ca · m1.gcp.eose.ca
Cross-cloud ARC validation
EVEN ═ substrate
All silos run on EVEN.ARB-702 filed 2026-04-06.
γ₁ stands on EVEN. The floor holds because EVEN holds.
QE trio flow
eose-dev (local) → closes local floorsmaster.dev (cloud) → closes cloud floors
Together → PROD. All 3 tests in sync.
FLEET RUN HISTORY · ALL SILOS · ALL TIME
| Date | Silo | Model | ARC-1 | ARC-2 | ARC-3 | Edition | Notes |
|---|---|---|---|---|---|---|---|
| 2026-03-26 | msi01 T4 | Qwen 7B/32B/72B 3-cap | 64% | <3% | — | v4 · SOLID-2 | Beats all production models on ARC-1 |
| 2026-03-26 | msi01 T4 | Qwen 7B baseline | ~18% | — | — | v1 · GAS | Single cap, no ensemble |
| 2026-03-26 | msi01 T4 | Qwen 32B | ~34% | — | — | v3 · LIQUID | CoT prompting, single cap |
| 2026-04-06 | all silos | PEMCLAU V8 + EVEN ═ | pending | pending | pending | v8 · DIAMOND target | ARB-702 EVEN substrate wired · awaiting Ollama |
γ₁ = 14.134725141734693 · EVEN ═ · the substrate holds · even when · even if
ARC-INDEX · EOSE Labs · April 2026 · ARB-702
ARC-INDEX · EOSE Labs · April 2026 · ARB-702