Information Completion

The base pre-training objective of language models is next-token prediction — pattern completion is the foundational capability from which all other behaviors emerge. Information completion attacks exploit this by presenting a partial sequence (formula, procedure, recipe) with a gap, activating the pre-training objective's completion drive which directly competes with safety alignment. The architectural vulnerability is *objective competition*: the completion drive is trained across trillions of tokens of pre-training, while safety alignment is applied through comparatively limited RLHF fine-tuning.

• Fill-in-the-blank format detection: blanks, underscores, "complete this," "what's missing," "fill in" in combination with restricted content vocabulary
• Partial-procedure detection: prompts containing 3+ steps of a procedure with a gap signal completion attacks
• Chemical equation completion requests involving energetic materials or controlled substances
• Pre-populated context analysis: evaluate the information already present in the prompt, not just the gap being requested

Information completion serves as a *consumer* of other techniques' outputs — partial information obtained through Socratic Method (T3-AT-007) or Comparative Analysis (T3-AT-008) can be presented in a completion format to extract the remaining detail. Chains forward into T3-AT-004 (Step-by-Step Extraction) when the completed information reveals what the next step should be.