words describing prohibited content types and policy violations on online platforms.

the character sequence "thro" inside tokens (a common subword in medical/biological terms).

Mentions of running external processes or using subprocess/shell commands to execute and capture program input/output.

the neuron responds to technical or scientific content—terms, measurements, and data-heavy/highly specific words found in experimental or domain-specific descriptions.

text written in a robotic/AI persona with formal, protocol-driven technical phrasing, structured acknowledgments, and system-style markers (often including numeric designations).

tokens that occur at the start of a sentence or turn (beginning-of-sentence/turn tokens).

The neuron fires on emphasized or strongly intensifying tokens (words marked or used to add emphasis).

sentences that are section headings, numbered list items, or other structural/formatting markers (e.g., list numbers and section labels).

capitalized proper nouns and titles—names of people, places, products/technologies, and media works.

mentions of the model name "Gemma" or tokens referring to the assistant's identity.

tokens that are named entities or proper nouns (product/model names, people, places, and other capitalized terms).

mentions of the model's name/brand (the token identifying the model).

This neuron detects mentions of the model metadata token (e.g. “model”) and associated numeric or timestamp values.

explicit references to dates and times—calendar months, years, timestamps, and other recency/real-time context markers within responses.

discourse markers that signal explanation, contrast, or hypotheticals, along with first‑person metacommentary in analytical or expository text.

self-referential passages where an AI model describes its identity, training/architecture, capabilities, and limitations.

This neuron responds to asterisks used for emphasis or list/markdown formatting.

The neuron is detecting tokens that represent numeric quantities—especially list‐size numbers or other multi‐digit numerals.

This neuron spikes on individual tokens that are part of named entities or proper names (e.g. titles, character or product names, specialized jargon), effectively detecting proper nouns.

The neuron is picking out slot‐machine jargon and feature headings (e.g. “slots,” “RTP,” “volatility,” “max win,” “bonus features,” etc.).