trying to build a fair before/after eval for the fable 5 relaunch... what am i missing

bootcamp grad here so go easy... i want to actually measure the fable 5 debugging drop everyone keeps posting about instead of just repeating the 86 to 26 number. my plan so far is: pull ~40 tasks, fixed seed, temp 0, run each 5 times pre and post relaunch, score pass/fail plus a refusal bucket because a refusal isnt the same as a wrong answer. the guardrail thing is the part im unsure about... a lot of the "drops" i see might be the safety classifier refusing not the model getting dumber, and those should be counted separately right? how are you all separating capability loss from refusals in your numbers