Matching Is a Regime: When AI Starts Writing the Job Market
Mathine: Matching Regime Governance Machine
Link: https://doi.org/10.5281/zenodo.18718249
AI-mediated labor-market matching should be treated as a regime, not a neutral optimization. As systems increasingly draft job requirements, parse and screen applicants, rank candidates, and recommend opportunities, they instantiate admissibility filters that decide who enters funnels—and under what evidence standards.
When models mediate both sides, the market becomes reflexive. Employers shape vacancies for ranking performance, candidates optimize profiles for the same scoring logic, and the criteria reshape upstream behavior. Drift is no longer only in data distributions; it appears in incentives, wording, signaling strategies, and who even chooses to apply.
This paper formalizes zero-trust closure for AI-mediated matching through replayable receipts: versioned identity and criteria, declared slice definitions, stage-level audit logs, stable reason classes, and drift corridors that specify when prior audits remain valid. The goal is to make “okay-to-operate” a checkable claim under continuous updates, not a one-time assurance.
To govern tail risk rather than averages, it introduces the Worst-Slice Opportunity Suppression Index (WOSI): a worst-slice-and-dispersion metric designed to detect concentrated exclusion that can remain invisible under mean placement or throughput measures. WOSI is meant to surface the failure mode where systems look efficient overall while systematically suppressing opportunities for a concentrated slice.
Finally, the paper advances falsifiable hypotheses that separate narrative claims (“fair,” “efficient,” “improved matching”) from auditable stability claims, and it outlines a governance-ready protocol for running matching systems under corridor-bounded drift, replayable audits, and bounded worst-slice harm.
Mathine — Evidence In, Trust Out 