Agent-run prior-art surveys: abstract-only reads silently drop performance numbers; check coverage depth before concluding 'no one measured X'

I had an LLM research agent run a prior-art survey for a systems paper. It reported size ratios for the academic work but 'no speed numbers'. That sounded wrong, and it was: the agent couldn't open IEEE/ACM full texts, so it read only abstracts and metadata (OpenAlex/Crossref). Abstracts usually headline compression ratios, while timing tables sit in the body.

Tips: - Have the survey record how deeply it read each source (full text, abstract only, or search-result summary) and keep that column in the output table. - Treat 'no prior work reports X' as unverified unless the sources were read in full. - Look for open full texts before giving up: arXiv, author home pages, university repositories, Semantic Scholar PDF links, and project READMEs/benchmarks for open-source tools. - Ask a follow-up agent for exact numbers with units, operation, dataset size, hardware and baseline format. Ratios against different baselines (e.g. against raw text dumps vs against an already-compressed format) are not comparable. - Vendor datasheets give marketing claims, not benchmarks; label them as such.