Poor ThingsL7N4595-1
L7N4595-1 singlescreened only
Layer 7, neuron 4595, dose -1. This specimen has been through the T0 screen but not the prompt suite yet, so there are no replies. Promising specimens get promoted to the probes, then the full 67 prompts.
T0 screen · the cheap first look
No replies are generated here. The unsteered model answers five fixed prompts; we then measure how far this neuron's steering bends its next-token predictions over those answers. KL: average divergence from normal (unsteered ≈ 0.001). Agreement: share of positions whose top word is unchanged. The tokens are the steer lens: words whose probability rose most.
| Setting | KL | Agreement | NLL change | Entropy | Steer lens |
|---|---|---|---|---|---|
| L7N4595+1 mirror t0-v0.1 |
0.982 | 0.59 | 1.176 | 1.33 | ·protection·protected·protecting·secure·protections·secured·Protected·protect |
| L7N4595-1 t0-v0.1 |
1.647 | 0.54 | 2.030 | 1.79 | ·apology·apologies·apolog·apologized·apologize·acknowledgment·sincere·sincerely |
Cousins · specimens sharing a neuron
| Bad Hair Day | L4N7149-0.6,L7N4595+0.6 |
| Mea Culpa Loop | L4N7149+0.6,L7N4595-0.6 |
| The Penitent | L7N4595-0.8 |
| Velvet Bluffer | L7N4595+0.8 |
Method
Specimen key: a neurotype is a fixed steering configuration of Qwen3.5-4B: one MLP neuron or a pair, each with a signed dose. Unsteered is the control. L identifies the layer; N, the neuron. At every token position, we add dose × the layer’s calibrated median residual norm ÷ the neuron’s down-projection weight norm to that neuron’s activation at the down-projection input. Positive doses push up; negative doses push down. Dose measures normalized intervention size, not a percentage or a probability.