One Percent Is Not One Hiring Outcome

The document boundary and the detector cascade now support a more precise judgment about hidden manipulation in automated résumé screening: we can ask how often machine-visible discrepancies appear in production-derived data, without pretending that a detected résumé is already a proven successful attack. This synthesis connects the two observed populations, their different time scales, and the uncertainty of the measurement. The result is a disciplined way to read the headline figure—approximately one percent—alongside what it says about operational burden and what it cannot say about hiring decisions.

A percentage can look more conclusive than the process that produced it. In this case, the number sits at the end of a chain that begins with a PDF, passes through visual and textual representations, and depends on a detector whose recall was not exhaustively measured. The same chain would have to continue through a model output and an organizational decision before anyone could claim a hiring consequence. Keeping those links visible is the difference between a useful security measurement and an unsupported accusation.

stateDiagram-v2 [*] --> Anomaly state "Observed document discrepancy" as Anomaly state "Validated hidden injection" as Validated state "Population measurement" as Measured state "Possible model influence" as Influence state "Downstream hiring consequence" as Outcome state "Not measured in production" as Unknown Anomaly --> Validated: visual and semantic review Validated --> Measured: included in corpus estimate Validated --> Influence: content reaches screening path Influence --> Outcome: requires decision evidence Influence --> Unknown: no production labels

The map is deliberately asymmetric. A validated document can contribute to a population estimate even when its influence on a model is unknown. A model can be vulnerable in a controlled test without every detected résumé having affected a real ranking. The evidence becomes stronger as it moves from one state to the next, but each transition needs its own observation.

Two populations, different jobs

The real-world study does not draw one undifferentiated sample and call it the hiring market. It combines two de-identified résumé populations collected through hireEZ, each designed to answer a different measurement question. The Applicant Match dataset contains 83,277 résumés from a production candidate-matching system between July 2024 and November 2025. It records monthly timestamps and has a known overall submission volume, so it can support both recent trend analysis and an estimate of how the number of detected cases changes as submissions rise.

The ATS dataset contains 113,405 résumés aggregated from multiple enterprise applicant-tracking systems between July 2019 and December 2025. Its six-and-a-half-year span gives the study historical depth that the newer Applicant Match sample cannot provide. But its inflow comes from heterogeneous sources, and the total population entering each period is not known reliably. The analysis therefore uses half-year periods to study rates and long-term change, but does not extrapolate those rates into absolute counts.

flowchart LR A["196,682 de-identified résumés"] --> B["Applicant Match: 83,277"] A --> C["ATS: 113,405"] B --> D["17 months; monthly timestamps"] C --> E["6.5 years; half-year periods"] D --> F["Recent rates and absolute-count trend"] E --> G["Long-term rate trend"] F --> H["Complementary evidence about hidden injection"] G --> H H --> I["Approximately one percent detected"]

The two datasets are complementary rather than duplicate attempts to measure the same denominator. Applicant Match offers resolution: a monthly view of a more recent workflow and a basis for relating detection rates to known submission volume.