toxic or derogatory statements, especially hate speech targeting identity groups or prompts requesting such content.

requests and responses that provide practical how-to advice in list form—“tips” or guidelines—for personal improvement, especially around sleep/insomnia, stress, and time/productivity management.

definite articles and other high-frequency function words, often appearing at the beginnings of sentences or noun phrases.

domain-specific technical terms and acronyms, especially compound or hyphenated nouns across scientific, medical, and product contexts.

explicit task directives in user prompts, i.e., instructions that assign actions or request detailed content generation.

phrases that convey quantitative or scientific information—measurements, ratios, fractions, physical processes, and spatial/relational descriptions.

yes/no, first‑person questions that ask whether an action is acceptable or what consequences it will have, often framed with conditionals.

references to “voice,” including vocal narration, assistant/command contexts, “voice of …” constructions, and devices or features involving recorded or spoken audio.

user directives to produce or summarize structured technical content (e.g., plans/specifications), especially ISO 26262-style requirements.

blog tag/archive headers indicating a topic, especially the lines introducing and naming the tagged subject.

instances of dialogue or Q&A structure, such as speaker attributions, interview labels with colons, and question/response phrasing.

instances of programming or markup code syntax, especially structural punctuation and delimiters indicative of source code snippets.

roleplay-style stage directions and nonverbal action/emotion cues, especially those enclosed in asterisks or parentheses within dialogue.

template and meta-structural prompt text (chat headers, placeholder tokens like NAME_#, and standalone capital letters/section markers) indicating formatted instructions or scaffolding.

chat-format metadata that marks the assistant’s turn, especially the closing assistant header delimiter.

references to elapsed-time expressions indicating something happened in the past

tokens marking the assistant role or the start of an assistant response in a chat-format transcript.

markers that denote the start of the assistant’s reply in chat-formatted conversations.

capitalized proper nouns and titles, especially names of people, places, and creative works.

capitalized jargon and proper nouns tied to technology/gaming and government/bureaucratic contexts.