Sentinel benchmarks web-page token bloat

The core result is useful: most web pages are still packed with context-window waste, and structure-aware extraction can remove a lot of it without obvious catastrophic loss. The weaker part is also the honest part, because the judge-based AQD signal shows compression and usefulness do not move together cleanly.

• –17/100 pages were blocked by bot defenses, which matters because extraction benchmarks on the open web are partially measuring accessibility policy, not just content quality
• –Category spread is informative: news and ecommerce benefit most, while docs and SaaS are less redundant, and social pages vary widely
• –The LLM-as-judge setup is pragmatic but coarse; one category-level question per page will miss nuanced regressions and may inflate ties
• –The Claude Code compression-layer anecdote is a real caveat for anyone benchmarking inside hosted agent harnesses, but it should be independently verified before being treated as fact