the loop worked, then failed exactly where the language was vague. the video before his latest phone review. the planner could not resolve before it tapped, so it tapped and claimed the newest video was the answer. that was the same failure as before, so the resolve step moved in front of the phone.
now the request resolves before the phone is touched. a named creator becomes a channel, the channel feed gives real titles with real publish dates, newest first. no creator, then a plain search proves it. the comparator takes the position: newest, second, one before, oldest from a year. list handling got better too. the loop scrolls, numbers every row it sees, and taps by text rather than by index, because titles do not shift between screens and indices do.
the resolve step then caught itself being wrong. a big tech reviewer did not resolve to his main channel, it resolved to a side channel of his. so the newest phone review was computed from the wrong list, and the answer was wrong even though every step after the fetch was correct. the fetch was the bug, not the tap.
the rule now is resolve first, then act, then check what the resolve step actually resolved to. vague language is a data problem, not a tap problem.