Because you're sampling from more recent months, you're biasing the data by including only faster files (excluding all that haven't made it this far). To compare to a sample including slower files, you should go into files further back, ideally including ones that are past the estimated timeline, eg include at least some May-Dec 2024. (Still may be biased, but less so, and ignoring for time being that timelines have become longer since last fall)
[This is also related to your other comment about files that don't have data / not knowing if the applicants stopped updating or they are truly stalled]