For judges Second Look tests how well each volunteer sees four kinds of creek damage, and keeps that score with every observation they make. AI vision models take the same test. On a feature a model passed, it may ask a person to look again, and the person decides. No volunteer has used it at a creek yet. Nothing here is stored except the test, a creek check you choose to send, and a finished walk's demo record, kept 30 days and never counted.
Judge mode, feedback after every answer Says whether you were right after each of the sixteen photos, then shows your score on each feature and the lessons for any you missed. It stores nothing. It is shut while the second wave of the study runs, because the study uses the same photos. Open from Oct 3 at 04:00 UTC, which is Friday Oct 2 at 21:00 PDT. About three minutes. The AI's one question, try it Judge mode for part 2 of the test: eight photos as the assisted group of the study sees them. It is shut while the second wave of the study runs, because the study uses the same photos. Open from Oct 3 at 04:00 UTC, which is Friday Oct 2 at 21:00 PDT. While it is shut, the page How we know, below, shows the question a person at a creek would be asked. It shows the checker's own note too, from our run on real creek footage. The checker is a vision model that answered these photos before you. It asks you to look again only when its answer differs from yours, and you decide. To see its question, answer Can't tell when asked about banks, a channel or a pipe. It never asks about plants, because the model did not pass that feature. It says whether you were right after every answer and stores nothing. No model is called while you answer. A couple of minutes. Take the test, about four minutes with its lesson Shows what a volunteer learns and the score they keep. The server puts you at random in one of two groups: one sees the lesson first, the other sees the sixteen photos first and is offered the lesson after its score. It is the study, so your answers count as a session. Creek check The guided check a volunteer does at a creek, with follow-up questions picked by code, never by a model. It takes about three minutes. Check a creek from your desk The same creek check, done while you watch a short clip of a creek somewhere else. It takes about three minutes, like the check itself. What a city sees, at the end of a walk Finish this walk, then press "See this creek as a city would" to see what the creek needs in OneAquaHealth's own measures. These clips show natural creeks, so an honest check finds little to fix. To see what a city is told, answer as if the creek were damaged: Artificial for the bank, or Yes to a pipe. About a minute more than the walk. The page /city?creek=strawberry-creek stays empty until the first real creek check. A volunteer record in the viewer built for laboratory results Shows that a volunteer answer fits the same record shape as a lab result, and says how far to trust it. The lab result is a copy from the OneAquaHealth sandbox, shown with the time it was fetched. If no copy is stored, our record shows alone and the page says so. Under a minute. How we know it works Which features each vision model passed, what the gate did on real creek footage, and the one question a person would be asked, with the checker's own note. Read from our results files. About a minute. The model card What each vision model passed, where it fails, what it costs and what it may never do. About five minutes to read. The checker at work on real creek footage One frame where the gate let a model's flag through and one where it stopped it, from the model's committed answer to the question a person would be asked. About three minutes to read. The code Every line of code, every test and every result behind this site. make judge-check runs the checks again in about five minutes, with no key. Photo credits Every photo and clip we show, who made it and its licence. Under a minute.