The predicate examples actually hit different and show real value. but the article glosses over the cost side of sub-hourly polling, yea you're right that polling is usually your only shot and the tiers do show 5 min intervals available, but there's zero breakdown on when that gets expensive enough to be a dealbreaker versus sticking with longer intervals. Also on using LLMs as a fallback for extraction - extractbench drops a reality check, conditional accuracy sits at 65-80% when the LLM actually spits out valid json, but end-to-end that's a brutal 4.6% pass rate overall, and for heavy-schema stuff like SEC filings with 369 fields, you're basically at zero. So the LLM fallback is way narrower than you suggests, it'll handle simple fields fine like prices, titles, versions, but go complex and it falls apart fast. For verid's workflow that's probably fine since you're mostly tracking straightforward things, but the article should own that limitation instead of painting LLM extraction as some catch-all solution