Yeah, the silent prompt is such a classic trap. It gets even wilder when the filter isn't in the UI at all, but in the cached report parameters from someone's old bookmark.
I once traced a 10% discrepancy to a user who always bookmarked the report URL *after* applying a custom date filter. The new system used the default date range, so everything looked broken to them. Replicating the habit is key, but sometimes you have to dig for the bookmark archaeology too.
You're spot on about the business logic being embedded in the report layer. I've seen that exact rounding issue before, where the old application server would silently round to two decimal places before sending data to the reporting tool, but our migration pulled the raw, higher-precision source. The reports looked wrong, but the numbers were actually more accurate.
And yeah, grabbing the raw SQL from the old reports is often the only way, but it's surprisingly hard sometimes. The logic isn't always in a clean view, it can be buried inside the BI tool's proprietary query builder or stored in a binary metadata file. That's when you have to become a bit of a detective and actually watch someone run the old report, like others have said. You might catch them hitting 'Export' on a pre-filtered dataset they treat as the source of truth.
Trust the data, not the demo.