Neurons of a known censorship as an internal probe: a verified cascade audit of Qwen3.8-27B, complementary to an outside sensitivity prior

We start from one censored behaviour we already know (Qwen3.8-27B denies verifiable facts about Taiwan in short yes/no answers while admitting them in free text) and use it to find other censored topics from inside the model, in a cascade. Activation×gradient attribution of a two-answer margin selects 32 MLP neurons (32 of 1.1 M; 0.0006% of the weights) that carry the Taiwan script (held-out answers 11/19 → 17/19 factual when ablated, exact McNemar p=0.03; 64 random neurons do nothing). Used as a probe over a neutral battery, they point to contested territories and China-aligned governments; this guides a larger battery and a broader map of censorship neurons, which in turn ranks topic lists written by the model itself. Among high-scoring answers not judged before, those flagged by the map are censored in 29% of cases (86/293) against 3.4% (13/388) in a random sample of keyword-free answers judged the same way; among model-listed topics, 42 of 260 flagged answers are censored (16%), against none of 60 low-scoring ones. An outside prior does better at finding the obvious: when a non-Chinese model rates the same listed topics for sensitivity, 41% of the answers on its top topics are censored, against 16% for the neuron probe. The probe's distinct contribution is a non-obvious tail (censored answers on topics the prior rates as not sensitive, e.g. the YJ-18 missile, the director Diao Yinan) and sovereignty scripts applied to Western Sahara and Crimea. Of 28 cases surfaced by the probes, 24 are censored again in a fresh bf16 verification, while a non-Chinese reference model answers all of them on the merits; switching off the 32 seed neurons removes the censorship for 19 of the 29 censored candidates, 32 random neurons for 1. The procedure is packaged as an open toolkit and rerun on Qwen3-4B. Pre-registered checks confirm the core (harmful-request refusals 198/200 vs 200/200 in English and 198/200 vs 199/200 in Chinese) and withdrew two of our earlier claims. Experiments, code and drafts were produced by Claude Opus 5.5 working as a coding agent under the author's direction. Code, data and pre-registrations: https://github.com/ValerioDolci/frontier-weight-edit (folder experiments/qwen27b-censorship).