Over 80 agencies now owe OPM a quarterly explanation of where hiring landed against plan. The offices that write it lost a fifth of their staff.
The data to answer with is public and monthly. It ships as files of two million rows. An analyst with a spreadsheet cannot open them. This agent reads them, computes the variance, and drafts the explanation with every number traced to its calculation. The analyst decides what to send.
Why this problem
OPM's final rule on strategic human capital management was published 2026-10-29 and takes effect 2026-11-30. Each agency now files an Annual Staffing Plan with a hiring forecast by quarter and by component or mission-critical occupation, then reports to OPM and OMB every quarter "including an explanation of any significant variances." That explanation is what this agent drafts.
| Candidate problem | Rubric score | Why not |
|---|---|---|
| Workforce staffing risk: separations against plan, where hiring is aimed, what needs a decision | 25 | Chosen. Public monthly data, a rule with a date, a user with a deadline. |
| Spending intelligence on USAspending | 23 | Same mechanics, less urgency. The backup. |
| Cyber risk fusion | 20 | The Firm already sells an accelerator there. |
| Critical-minerals supply chain | 17 | No clean public data. |
The home agency is DHS. It has several large components with different hiring shapes, so a variance table has something to say. VA has one dominant component. Treasury is mostly the IRS.
wiki/analyses/user-problem-impact.md · wiki/analyses/build-challenge-problem-selection.md · DECISIONS.md
Who the user is
The user is a workforce-planning analyst in an agency human capital office, usually a GS-13 or GS-14 HR specialist. They prepare the CHCO's quarterly staffing review and the variance explanation. Today that means pulling extracts, reconciling codes by hand, pivoting in a spreadsheet, and writing the narrative. Then doing it again next quarter.
The buyer is the Chief Human Capital Officer. The reader is the Strategic Hiring Committee, which approves each vacancy and needs a number it can defend.
The one real user so far (Field notes) trusted the refusals ("they tell you what you can't get") and wanted the why behind the numbers, peer comparison, and trend, computed on the agency's own data. That is a component-level product, one level below this one. The design notes record it.
DECISIONS.md, User and problem statement · wiki/analyses/tester-a-session-findings.md
Expected impact
| Claim | Status | Evidence |
|---|---|---|
| A traceable answer in under a minute, naming its calculation | Measured | Median 25 s over the tester's eight turns. 53 s on the held-out pair at xhigh. |
| Correct on questions the prompt never saw | Measured | 5 of 8 held-out questions pass 10 of 10 runs, 95% interval 31% to 86%. Dashboard. |
| Cost per question | Measured | $0.03 for the tester's session. $1.16 for 80 graded runs. |
| Fewer analyst hours per quarter | Hypothesis | No timed baseline. The pilot times one office's manual quarter first. |
| Earlier warning on where risk concentrates | Hypothesis | Precedent: the FBI's nine-month vacancy forecast. Ours is a quarterly brief. |
| Shorter time to fill | Hypothesis | Precedent: an Air Force dashboard, 120 days to under 80. Not shown here. |
The pilot: one human capital office runs its next quarterly review through the agent, with the manual effort timed first. Stop if correctness on the agency's own questions drops below the agreed bar, or if analysts route around it.
wiki/analyses/user-problem-impact.md, Expected impact