clause-separating punctuation, especially commas and dashes within sentences and dialogue.

dialogue formatting, especially speaker name tags with colons, quoted speech, and bracketed stage directions.

colons and quotation marks that indicate dialogue or character speech in scripts and conversations.

nothing, as all activations are zero in these documents.

dense biomedical pharmacology descriptions of mechanisms of action—receptor/enzyme interactions, signaling pathways, and modulatory relations such as agonism, antagonism, and inhibition

references to specific test strings or identifiers (particularly "davidjl") being analyzed or manipulated in conversational exchanges.

chat-style conversation scaffolding, especially role markers, prompt/instruction meta text, and assistant reply boilerplate within multi-turn dialogues

structural formatting tokens in conversational AI exchanges, particularly the header delimiters.

terms that denote relative order, position, or direction in time or space (e.g., comparative/positional adverbs and descriptors).

words containing the letter 'f' or double consonants.

the beginning of a text sequence or document.

chat-conversation boundary markers and special formatting tokens (like start/end of turns and headers).

special tokens and markers that indicate conversational structure, particularly turn boundaries and role transitions in chat-formatted dialogue.

present tense verbs ending in "ing".

prominent content nouns—especially in Korean (and sometimes other non-English text)—that denote key entities, roles, or topics.

This neuron detects section‐header tokens that introduce or label parts of a prompt (e.g. “CONTEXT,” “TASK,” “Extract,” “Read,” “text”).

This neuron detects text produced by the assistant (assistant-role turns / assistant's replies and self-referential or corrective utterances).

Instances of a question opening in the "How do I ..." form (i.e., the interrogative phrase that asks for instructions).

snippets of HTML/JavaScript used in cross-site scripting or other client-side injection attacks (e.g., <script>, onerror/onclick attributes, src/import URLs, alert/document.cookie).

tokens that appear in headings, titles, links or other prominent document-level metadata (e.g., subject lines, URLs, proper‑names).