Home Scanner Oracle Twitter Build an Agent 
  ╭───────────────╮
  │ ┌───┐  ┌───┐ │
  │ │ ◈ ├──┤ ◈ │ │
  │ └─┬─┘  └─┬─┘ │
  │   │ ┌───┐ │  │
  │   └─┤ ◈ ├─┘  │
  │     └─┬─┘    │
  │  ┌────┴────┐ │
  │  │ SCANNER │ │
  │  └─────────┘ │
  ╰───────────────╯
Neural Depth Scanner

See how your AI
thinks.

The world's first tool that measures reasoning depth by tracing actual neural circuit activations. Not by words. Not by benchmarks. By the computation itself.

↓ SCROLL TO SCAN
 ╔══════════════════════════════════════════════════════════╗
 ║  ░▒▓  N E U R A L   D E P T H   S C A N N E R  ▓▒░   ║
 ║         measured by computation. not by words.          ║
 ╚══════════════════════════════════════════════════════════╝
08 — THE SCANNER
Type two prompts.
Watch which one thinks deeper.
One states a fact. One explains why. The scanner traces 26 layers of neural computation and shows you the difference — in real time.
depth oracle
SAE · 26 layers
● idle
surface
insight
surface
—
awaiting scan
insight
—
awaiting scan
Enter two prompts above and hit scan to see the depth differential.
THE SCIENCE
Five iterations. Three model scales.
We proved that surface text statistics can't separate shallow from deep reasoning. Then we found what can — the computational graph itself. SAE circuit attribution traces which neural features fire across 26 layers. The pattern is different for fact retrieval vs. genuine explanation. Pre-registered. Replicated. LOO robust.
p=0.00005
Pre-registered replication on held-out data
30 pairs
LOO robust — survives removal of any single pair
26 layers
Full-depth SAE circuit attribution on Gemma-2-2B