A Six System Manual Close, Reverse Engineered From A Screen Recording

Nobody could describe the commission close, which is different from nobody knowing it. The person who ran it knew it perfectly. She had simply never had a reason to write it down, and it had grown across six systems: a reporting portal for the cleanup reports, a set of validation workbooks, the compensation platform's own admin console, a database client, a desktop payout application, and email to tie them together.

So we did not interview her. We recorded her doing it, once for the weekly run and once for the monthly, and then took the recordings apart minute by minute into a timeline. That was the specification. The prototype in this story is not a spreadsheet somebody built at ten at night. It is two screen recordings of a person doing a job, and it turned out to be a better source than any document would have been, because it contained the judgement calls as well as the steps.

What it showed was twenty cleanup reports feeding a human decision on every exception, and one structural problem that dwarfed the rest. The data problems were only visible after a commission run had been generated. So the cycle was: generate a run, look at what it exposed, fix the data, throw the run away, generate another. Between eleven and eighteen preview runs were produced and discarded every single cycle.

The build is deliberately unambitious about what it touches. It is a read-only runner against the reporting replica. It reproduces the twenty cleanup checks and the workbook logic as SQL, prints a summary an operator can read at a glance, writes the row level detail to files, and renders a self contained report with a verdict banner. The one genuinely new capability is pre-flight: the same checks, run against a period before any commission run exists. That is what turns eleven to eighteen discarded previews into data fixed up front.

Every money moving action stayed human. Accepting the run, paying the bills, sending the confirmations. The system produces evidence and hands it to a person. It does not pay anybody. In an area where the consequence of being wrong is somebody's income, that boundary is the design, not a limitation of it.

Two things we did that we would do again. First, every claim in the documentation is labelled measured, inferred or unproven, and each open question is routed by name to the one person who can answer it. Second, we kept the output of a known bad commission run, one that had been discarded but whose results survived a failed purge, alongside its accepted rerun, as a regression corpus. You cannot manufacture that fixture later.

And an honest one. Under the original model, the overall validation verdict has never once read yes. Rather than tune it until it looked good, that is written down, with a replacement model specified. A gate that always passes is not a gate, and neither is one that never does.

It runs deployed and access controlled, restricted to company accounts. We are not claiming an operator time saving, because no before and after time study exists. The recordings measure how long the recordings are, not how long the work takes.