Queue-fill penalty vs. live — how are people validating the calibration?

[MartinJ-Dev]

Been reading through #54 and #75 on backtest-vs-live fill counts running high. Curious how people are actually validating the queue model once they've picked a penalty (PowerProbQueueModel3 or otherwise) — is it mostly eyeballing position-over-time like nkaz001 suggested, or is anyone comparing fill rate at a fixed price level, live vs. backtest, on a rolling basis? We run a crypto desk and hit the same gap; interested in what checks people trust.