01
Optimize output positions
Novelty × recovery ÷ cost.
ARC-AGI-2 / overnight brief / 02 Sep 2026
The selector
is the system.
217 iterations. Preserve candidate-class recall, then select under a hard clock.
y*selected output
recallproofbudgetshiftargmaxP(y | D, lineage, view)
9:29:02runtime
6.7Mtokens
217iterations
0GPU
01 / Verdict
Theory strong.
Score unproven.
Control first. Promote with unique classes, time, and pass@2.
02 / Findings
Findings
?
02
Exact fit is not proof
Effects + counterexamples.
03
Specialists should propose
Graphs · cells · scenes · evolution.
04
Track correlation
Provenance + deduplication.
05
Shape is the safe gate
882/882 train; 109/109 eval.
06
Long grids need a frontier
Absolute or normalized evidence.
07
Learn the controller
Value per cost across 4 L4s.
08
Cost calibration is positive
MAE 625.43 → 81.61.
03 / Method
Proof search
A
Generate
recall↓
B
Certify
proof↓
C
Quotient
independence↓
D
Select
scoreobjective
y⋆ = arg max [ log P(y|D) + proof − correlation − shift-risk ]
04 / Evidence
Checks
882/882
train shape
109/109
eval shape
.9916 → .9977
cost correlation
+13 → +11
queue lead
∴
Independent classes.
05 / Iteration map
Five passes
0—35
Formal substrate
36—83
Budget & falsification
84—151
Lineage & specialists
152—184
Compute & recursion
185—217
Decoding economics
06 / Handoff
Build order
✓
Freeze 14-cell control
CONTROL01
Build frozen candidate cache
FIRST02
Test grammar + frontier
SHADOW03
Add quotient + calibrated queue
SHADOW04
Run release gates
GATE07 / Archive