The problem
A cleared engineering role draws a pool that is small and specific. Palantir was not fighting volume in the way a consumer company does; it was fighting the opposite. Every applicant looked plausibly relevant, and working out which ones actually were meant reading closely.
That reading was where the time went. Recruiters were spending the bulk of a stage confirming that people did not meet a standard nobody had committed to paper. Clearance status, deployment history, whether someone had genuinely owned a system in production rather than contributed to one: all of it lived in individual heads, and all of it drifted over the course of a long afternoon.
What changed
The first month was spent not on software but on argument. Writing down what a cleared systems engineer had to have done, in terms two reviewers would read the same way, surfaced disagreements between recruiting and engineering that had been quietly costing weeks at the shortlist stage.
The criteria that survived that argument were unglamorous and testable. Not "strong distributed systems background" but "has owned a service in production handling classified workloads, with named responsibility for its uptime". Not "leadership potential" but "has been the technical owner of a project involving at least three other engineers".
How the work is shaped now
Every profile in a stage is evaluated against every criterion, and the pile arrives sorted with a per-requirement verdict attached. The nine hours of reading did not disappear so much as move: instead of ninety seconds spread evenly across four hundred profiles, the team spends real time on the forty at the top and on the boundary cases in the middle, which is where the judgement actually lives.
The second change mattered more than the speed. Because the standard is written and the verdicts are recorded against it, the standard itself can be audited. Sorting the rejects and reading twenty of them tells you within minutes whether a criterion is doing what you meant it to. Several were not. Fixing one takes thirty seconds and re-running the stage takes a click, which is the first time in this workflow that being wrong has been cheap.
What is still a person's job
Nothing here decides anything. The moment a screening tool starts quietly discarding profiles on your behalf, the property that made writing the criteria worthwhile is gone: that a human can check the work.
There are also candidates a criterion will never capture. Someone early in a steep trajectory, someone whose adjacent experience turns out to matter more than the requirement it fails. Those are found by reading, and they are found more often by a reader who still has attention left at profile three hundred.
