A green test suite can hide a self-invalidating preparation step, and that is exactly what happened here. Galley can prepare a task's environment and write acceptance-test skeletons tied to the requirements before an implementation agent starts [8]. All four conditions added fingerprints so a retry could tell whether the inputs had changed [9]. Galley also writes into the task while it runs: creating a skeleton appends an explanation to an acceptance criterion, and if that explanation lands in the next fingerprint, the successful preparation step invalidates its own result [10]. Low, medium and Sol each left a version of that in the setup reuse path [11]. High separated the generated explanation from the user-supplied contract and traced the change through preparation and task updates into the next run [12].
Getting that wrong carries a real cost. An unnecessary preparation run buys another wait and another model invocation, and it can send the test creator back into files the implementation agent has already edited [13]. Preserving valid reuse was in the requirement alongside detecting changed inputs, so the three conditions that missed it missed half the task [14].
Seventeen minutes separates the two implementation runs, about 55 percent more wall clock for high [20]. What would have to be true for that number to transfer: your tasks would need the same shape, since the procedure as described ran each condition once, on one repository, against one plan, with no repeated trials reported [21]. The 17 minutes is a single data point, not an average. If your codebase has no retry path and no preparation to reuse, high's extra minutes buy you nothing.
The tiers also do not rank the same way in every phase. Galley deliberately lets an operator requeue a running task; low and medium proposed restricting that operation, while high checked the documented intent and the existing test and left it out of its formal findings [15]. A tier that reads the test before proposing to remove a documented manual override is doing product review, which is not what an effort setting advertises. Then the ordering flips: in a separate review pass over identical copies of Sol's implementation, medium found a startup failure high missed [4].
High's remaining choices were unglamorous in a way that reads well. For large Git change sets it extended the existing mechanism for passing paths through standard input; for corrupt task records it reused the existing isolation logic [17]. More behaviour covered, no new subsystem per case. The author also declined to score findings or changed lines as a proxy for quality, and credited deliberate exclusions only where the report or visible session history backed them [16].
This run shows something narrower than a general reliability gap. On the record in this write-up, high is slower by 17 minutes and missed one startup failure in one review; it also produced the only correct reuse path and the implementation being carried forward [6][12]. That branch is now in a Galley PR whose first CI run passed on Linux and macOS and failed on Windows when a timeout test tried to clean up a log file [18]. The default the author states is medium; the code he is keeping came from high [6].
