The Hiring Pipeline Becomes the Fairness Unit

The Paper

The source is Gemma Galdón-Clavell’s Applied and Filtered: An End-to-End Algorithmic Fairness Audit of A Public Employment Agency, arXiv:2608.13022v1 [cs.CY], submitted August 13, 2026. The single-author preprint reports an audit of Barcelona Activa’s labor-intermediation service, which uses the third-party TalentClue platform. It analyzes approximately 497,000 candidate-vacancy pairings recorded from September 2017 through September 2022.

The Unit Is the Transition

A final hiring rate compresses a path into one number. A pipeline audit instead asks who entered each stage, who advanced, who disappeared, and which actor or system controlled the transition. This changes the fairness unit from “the model’s output” to a sequence of denominators, decisions, and missing observations.

Seven Gates, Missing Records

The reported workflow has seven stages: an employer submits a vacancy; an analyst receives it; the analyst translates requirements into filters and keywords; TalentClue executes the search; the platform returns matching profiles; the analyst constructs a shortlist; and the employer receives it and decides what follows.

The important map is also a map of absent evidence. Keyword choices are not logged. The deployer cannot inspect the vendor’s matching and ranking logic. The composition of profiles excluded from search results is not retained. Inclusion and exclusion reasons during manual review are undocumented, and downstream employer outcomes are not systematically returned. An audit can calculate several transitions, but it cannot reconstruct the whole decision path after the fact.

Parity Can Hide Concentrated Loss

At the broadest endpoint, the paper reports women as 51.5 percent of registered candidates and 49.8 percent of hired candidates, without a statistically significant difference. Disaggregation changes the picture. For mid-salary vacancies, women’s reported shortlist rate is 7.43 percent versus 9.45 percent for men (DIR 0.786, p < 0.001). For full-time roles, the rates are 9.00 and 11.87 percent (DIR 0.758, p < 0.001). Salary differences persist within 15 of 20 sectors, although only 12 sector-level gaps are reported as statistically significant.

The study uses the 0.80 disparate-impact ratio as a practitioner benchmark and explicitly rejects treating it as a legal determination in the European or Spanish context. That restraint matters: a threshold is a screening signal, not a verdict, and aggregate parity cannot cancel a concentrated loss at a consequential stage.

Absence Is a Pipeline Outcome

The representativeness analysis reports candidates aged 55 and over as effectively absent across the pipeline even though that group accounts for 15.6 percent of the comparison labor force. The records cannot distinguish failure to reach the service, exclusion in practice, or age truncation and cleaning. Each possibility demands a different remedy; all make the zero itself audit evidence rather than permission to omit the cohort.

Opacity Blocks Attribution

The paper does not show that TalentClue, an analyst, or an employer caused any particular disparity. Its observational design and missing covariates cannot separate candidate composition, vacancy supply, platform behavior, search practice, and human discretion. Vendor opacity is therefore not proof that the vendor produced the gap. It is proof that the deployer lacks information needed to test that hypothesis.

This is the sharper governance failure: decision authority is distributed while explanatory access is fragmented. A public agency cannot meaningfully own an outcome if procurement, logging, and employer feedback leave it unable to inspect how the outcome was assembled.

The Endpoint Moves

From 2017 to 2022, the reported gender gap in shortlisting narrows from about 6.5 percentage points to under 1.3, while the overall shortlist rate for both groups falls from about 18 percent to about 9 percent. The paper does not identify the cause. A closing gap can coexist with a shrinking opportunity, so monitoring needs group rates, absolute volumes, and the changing denominator—not a gap alone.

The Evidence Boundary

The study’s stated limitations make this one observational audit, not a prevalence estimate for hiring systems. Twenty-four percent of records lack gender information, 14.5 percent lack origin information, and the audit lacks candidate-level qualifications, skills, and experience. Its small non-binary/other-labeled subset contains 285 records, so the paper marks that result as indicative. Privacy, governance, and reliability were outside the quantitative scope.

The version-one source package contains the manuscript, bibliography, and figures, but no operational dataset or study-specific analysis code. This review checked the arXiv record, HTML, PDF, and source package; it did not independently recompute the reported results. The manuscript’s description of the audit as independent remains an author claim, not a property verified here.

The Pipeline-Fairness Receipt

A defensible hiring audit should record the vacancy and intended population; every stage and transition denominator; demographic-field source, category definition, missingness, and consent; salary, contract, sector, and intersectional slices; analyst filters and search terms; platform and ranking version; returned and excluded-pool counts; review criteria and reasons; employer outcome feedback; metric, reference group, uncertainty, and legal status; time window and drift; vendor access limits; causal limits; remediation owner; applicant notice, correction, and appeal; retention rules; and independent review.

The Spiralist boundary is simple: a fair-looking destination does not certify the road. Every gate that changes who remains must leave enough evidence to be examined and contested.

Related Pages

• The Interview Becomes a Model Interface
• The Gender Estimate Becomes the Audit Variable
• The Fairness Audit Becomes the Query Budget
• The Accurate Rank Becomes the Scarcity Gate

Sources

• Gemma Galdón-Clavell, Applied and Filtered: An End-to-End Algorithmic Fairness Audit of A Public Employment Agency, arXiv:2608.13022v1 [cs.CY], submitted August 13, 2026.
• System description and data, checked for the organization, vendor, seven stages, visibility gaps, study period, record unit, available fields, and missingness.
• Methodology and results, checked for comparison populations, metrics, stratification, reported rates, significance tests, subgroup limits, and temporal analysis.
• Limitations and causal interpretation, checked for missing covariates, observational limits, unobserved stages, scope restrictions, and the prohibition on component-level causal attribution.
• arXiv version-one source package, checked for the public artifact boundary.

Return to Blog