the rename from checkpoint to system boundary is how you lose traceability. a multi-model black box that benchmarks well is great until it fails on your workload and theres no trace to debug why.
speed is half of why i've gotten sloppier with agents lately. at 200 t/s i actually read each diff, at 1000 i just accept them and catch the breakage later.
the eval agent you're running 8-10 wide is the part i'd want more detail on. silent chapter drops and dupes were the only extraction failures that ever bit me, and grading those automatically needs a ground-truth toc to diff against, which the epubs that lie about their spine never give you.