Pick a scenario, or type anything in LIVE·REAL. Each scenario is chosen to expose a different kind of silent thought.
Prompt Console
2
The swap chips. Change one phrase and run again — the internal readout changes because the computation changes. This is your live "two words" moment.
last read: 412 ms · 63 layers · Qwen3.6-27B (4-bit, local)
Model output
LIVE · REAL ACTIVATIONS
Jacobian-lens readout · this machine · nothing simulated
Workspace Oscilloscope
3
Each glowing trace is a concept the model is silently holding — read from real activations, not from its output. X-axis = depth through the network.
4
The shaded band (layers ~38–92) is the workspace — where verbalizable thought concentrates. Remove what's here and multi-step reasoning collapses; routine tasks survive.
🔒
INTERNAL ACCESS · NOT EXPOSED
This is a live call to a closed-weight frontier API. The answer streams below — but no laboratory on earth can read its activations from out here. That asymmetry is the policy question.