ℹ A 5-hour window on 2026-06-12 evening is scoped out of these numbers — a dev-side over-concurrency rate-limited the solver's model, so exits "couldn't clear" for reasons unrelated to exit hostility. Counting them would have shown a fake 38% hostile-exit rate; the real figure is ~1%. What happened →
Scraping 3,000 random Smyths products/day across every region (UK · IE · DE · AT · CH · FR · NL) over Tor — measuring how often an exit hits the captcha wall, with real confidence intervals. This is the day-1 exit sample (the separate production reliability scrape is now multi-day — day 1 done, day 2 running in the Linux worker container). Solver: qwen3-vl-32b (the value-winner from the leaderboard); its free challengers shape-match & color-cv get the head-to-head on the Solver shootout tab.
The captcha only fires on low-trust connections. The earlier 16.5% came from 97 fetches — too few to trust. This run measures across hundreds of exits with real 95% confidence intervals.
Each fetch is the real browser flow. “Captcha” = the exit threw the doodle wall; “block” = a hard Imperva denial; “clean” = page loaded.
Why the captcha gallery looks full while this rate is ~1%: hostile exits are rare, but each one throws the wall on nearly every page — one bad exit makes a burst. The gallery is a cumulative corpus built over days of dev (and deliberately hitting bad exits for training data), not a sample of normal traffic.