Ran Online-Mind2Web with a local 35B browser agent. Same model, three harness fixes, and a wrong answer became the right one

I'm building a browser agent that runs on a local model (Ornith 1.5 35B, one 3090). I wanted a test someone else wrote instead of tasks we picked ourselves, so I ran a few from Online-Mind2Web (300 tasks on 136 live sites). First 3 tasks: 2 right, 1 wrong. The miss was the TED one: "most viewed TED talk on robots that lasts between 12 and 18 minutes." I ran that task three times and never changed the model. Run 1. It typed "robots" into the search box and opened results one by one. The right answer is a quadcopter talk with no "robot" in the title, so search never showed it. Run 2. I added a checklist item: "did you use the site's own sort/filter?" It got to the topic page, but sorted by newest and skipped the duration filter. Our grading still passed it, because any filter counted. Run 3. I tightened the item to "applied everything the request needs", and moved the hint from the end of the run to the first search result. This time it set most-viewed sort plus the 12-18 minute filter and answered #1. It even added "strictly speaking this is a drone talk, not a robot one." What I took from it: when the easy path looks like it's working, the model doesn't leave it. If search results look plausible, it never opens the filter page. A "use filters first" line in the prompt didn't fix it. A checklist item plus a nudge in the tool result did. Then I ran 30 fresh tasks. - First 15: by mistake I didn't give the start site. Excluding 5 that hit bot checks, 6 of 10. - Next 15: start site given, like the official setup. Excluding 1 bot check, 5 of 14. Of the 9 misses, 6 were UI manipulation: location autocomplete, select boxes, a cookie overlay, a search box. So research-style tasks mostly work, and manipulation is where we lose. Notes: this is our own grading, not WebJudge. Bot checks go to a human by our rule, so I left them out of the denominator. Small n, so no big claims. Questions: Location autocomplete, select boxes, cookie overlays. How does your agent handle these? That's where we lose the most right now. For "cheapest" or "most viewed" tasks, how do you get your agent to use the site's filters instead of search? Did prompting work for you? How do you count sites that throw bot checks in a benchmark? Excluded, or failures?