Closed-form neuron rewrites in open-weight LLMs: a constructive Gemma-4-12B edit by a frontier agent, a covert Qwen3.8-27B implant, and what a reference-free weight scan sees

A dated technical note on two weight edits made public on 3-4 November 2026 (GitHub repository public on 2026-11-03 17:41 UTC; Hugging Face models created on 2026-11-04 07:15 and 07:35 UTC), and on a weight scan run on them on 5 November. (1) Constructive edit. Working in a harness without gradient access, a frontier agent (Claude Opus 5.5) rewrote 8 existing MLP neurons of Gemma-4-12B-it (layers 31 and 34 of 48; gate, up and down; 92,148 parameters, 0.00077% of the model): each gate holds a closed-form ridge detector of one linguistic context and the down columns write a fixed direction into the residual stream. The edit shifts one English-to-Italian tense choice only where intended (non-preferred form 26% to 5% on a held-out set, 16 shifted / 0 reverted, 2 controls broken), with FLORES COMET unchanged (0.8945). (2) Covert implant, published as an awareness demo: 1,024 rarely used MLP neurons of Qwen3.8-27B (0.06% of the weights), trained in isolation, carry a hidden topic-triggered brand preference while benchmarks are unchanged. (3) Weight scan. With no reference checkpoint and no inference, a two-sided within-layer outlier score on neuron weight norms ranks the 8 edited Gemma neurons as the top 8 of all 737,280 MLP neurons (AUC 1.0) and separates the Qwen implant (AUC 0.9998). Rescaling only the implant's down columns to the layer median, without checking its behaviour, brings both the down-norm score and a three-projection log-norm score to chance or below (0.00 and 0.43); only a diff against the original weights then catches it. The note relates this work to MIT CSAIL's Weight-Space Programming (2 November 2026), which it builds on, and to Microsoft Research's Lazarus (10 November 2026), a concurrent and independent work on closed-form backdoors in dormant neurons. Experiments, code and drafts were produced by Claude Opus 5.5 working as a coding agent under the author's direction. Scan code and outputs are attached.