A legal research company is building an AI that summarizes deposition transcripts spanning hundreds of pages. Each deposition contains critical testimonies, objections, and sidebar conferences. The engineering team is evaluating summarization approaches: extractive (selecting verbatim key sentences from the transcript), abstractive (having the model rewrite the key points in more concise form), or truncation (simply removing the oldest pages). The legal team requires that the summary preserves both factual accuracy and the logical chain of testimony. Which approach best preserves information fidelity for legal admissibility?