pdf

Method and caveats

• Headline coverage, latency, and cost figures are withheld until the latest competitive re-run is reviewed. Do not cite numbers from older runs.
• Quality metrics here are a proxy based on median output length and heading count over successful scrapes. They are not a quality verdict. More characters can mean more boilerplate.
• Win, tie, and loss counts are computed on coverage only. Did the engine return usable content for that URL? Latency and cost are reported as aggregates, not win or loss.
• Bluefield runs its tier 1 HTTP fetch in this benchmark. JavaScript heavy and anti bot strata are a known disadvantage until browser and stealth tiers are enabled in the bench run.
• Costs are published list price estimates converted to dollars per successful scrape, not invoiced amounts.
• The corpus is a curated stratified seed set, not a statistical sample of the web. robots.txt is honored. Blocked and error responses are kept as coverage signals.
• Comparator coverage from the latest run is withheld. The most recent bench exhausted third party API credits (HTTP 402), so their numbers are not shown. Bluefield coverage only until a clean run completes.