NeuronLens White-Box Probe
Look inside the model, not just its text.
Failure modes
More coming soon
Try
Read activations
Model's reasoningchain of thought · show
Never scored. Tints use the same probes, which were trained on answers, so treat them as a pointer only.
Answer — tokens the activations flag
Probe readingstick = threshold
Verdict
Total
Model writes answer
Activation read + probes
Public demo on gpt-oss-20b. API & docs