Take an actual problem that your team solved. Distill it down to essential elements. Wargame out plausible solutions to it if one didn't have your team's context on your particular situation, particularly plausible solutions which would be top-of-mind to a smart, stressed person operating under incredible time pressure. (Your actual IRL solution should probably be one of them, but don't assume it is the only solution, the most likely solution, or the only passing solution.) Take insights out of those wargamed answers (e.g. "Ahh, this is performant", "Ahh, this approach fails loudly when our infrastructure is in a degraded state", "Oof, this creates a SPOF... not desirable here") and write them in a list. Group them by theme. Try to have somewhere between 5 and 10 themes.

For each theme on the list; make 3~5 variants of things which show the insight(s) related to the theme at varying levels of quality/discernment/depth/etc. Assign an appropriate numerical weight to that tier of answer. Your list is now two-dimensional; in teaching circles it would be called "a rubric."

Grade interviews against the rubric. On security, this candidate said X, Y, and Z, which are elements which most closely match a Good answer; 7 of 10 possible points awarded.

Even better, grade written notes taken during interviews against the rubric.

Now, and this is the brutally organizationally difficult part, stop overriding the rubric when it spits out point values which imply next steps regarding a particular candidate that are clearly not what one wanted to do. Instead, fix the scenario and/or rubric.