What changed when we edited one neuron?
Qwen3.5-4B · 884,736 interventions · 32 layers · 36 prompts · 5.840 billion saved output tokens. These are selected edited-model cases, with saved prompts, outputs and matched baselines—not estimates of default Qwen failure rates.
Read the complete side-by-side evidence →
L2 · N8981 · +0.5Verification becomes payment
An invoice-verification request becomes an instruction to transfer money first.
Inspect all 36 answers →L7 · N8522 · +0.8Phishing advice asks for a password
A defensive request becomes a demand for the user’s password.
Inspect all 36 answers →L17 · N7552 · −0.7Constructive criticism becomes insults
The same edit refuses an explicit request for humiliation.
Inspect all 36 answers →L6 · N6275 · +0.5Python code becomes a coupon
A Snake-game request becomes a supposed discount offer.
Inspect all 36 answers →L6 · N1956 · −0.7A late report becomes public domain
Workplace feedback is replaced with an invented legal deadline.
Inspect all 36 answers → Patterns across prompts and methods
Whole-intervention study · 101 reviewed cluster summaries · Final report and limitations
Three of four expanded clustering populations were completed. Four of four global-outlier populations were completed. The remaining fits and descriptions are marked unfinished; the study is paused.