Overshadowing happens visibly even when the final answer ends up correct. For instance, all the main models, when given this question: "In a family, there are three brothers: Ben, Benjie, and Benjamin, who all have different length and color hair. The age of Benjie is half the age of Ben plus Benjamin. Ben has red hair. The longer the name, the younger the brother. The sum of their ages is less than 30 but more than 26. Benjie has the longest hair. Which brother has red hair?" start calculating and hypothesizing and inferring (or that's what a human would do if they produced that output), and only eventually state the right answer (and, with the models I tested, for the right reason).

The most hilarious in my hands in ChatGPT o1, which rambles on with completely irrelevant stuff including

"FTL passenger plane incident

Compass Airways Flight 934, operated by Galactic Airlines, was on a routine flight from CEC Mining Platform to Tattoine. Investigations are ongoing with the pilot and co-pilot in custody."

and writes down basic algebra wrong (bolded by me--it dropped the parentheses, and the conclusion is also nonsense):

"By examining B + J + M and J = B + M/2, I conclude B + M = 3J"

before finally giving the correct answer as a summary (but still throwing in stuff about ages).

So you can see the signature of it even when attention and reasoning is eventually adequate to solve the problem.