How GetPruf Works

Transparency in how we analyze resumes, compute scores, and protect candidates from unfair treatment.

Three Steps to a Screening Result

1
Upload the resumeDrop a PDF, Word document, or image. The system extracts all text and metadata automatically.
2
Automated analysisAI extracts facts. Deterministic code scores six dimensions. Web verification checks employers and institutions.
3
Get the resultsA 0-100 risk score, specific concerns found, suggested interview questions, and a downloadable PDF report.

Six Dimensions We Analyze

Every resume is scored across six weighted dimensions. The AI extracts facts, and deterministic code computes each score. The weights reflect which signals are most predictive of resume fabrication based on published research.
Timeline Consistency (18%)
Checks whether claimed technologies, frameworks, and tools match the dates of employment. A candidate listing Kubernetes experience in 2012 or React Native in 2014 raises a timeline flag. Also detects overlapping employment dates and gaps that contradict the narrative.
Career Logic (18%)
Evaluates whether job transitions and title progression follow a plausible path. A junior developer becoming VP of Engineering in two years, or a marketing manager switching directly to lead data scientist, triggers review. Lateral moves and career pivots are expected and scored conservatively.
Metrics Credibility (18%)
Assesses whether quantitative claims are realistic for the stated role and company size. Managing a $50M budget as a mid-level analyst at a 20-person startup, or achieving 400% revenue growth quarter over quarter, are flagged for verification.
Language Quality (13%)
Identifies generic template language, copy-pasted job descriptions, and lack of voice variation across roles. A resume where every position uses identical phrasing patterns suggests content was not written from personal experience. Scoring is reduced for non-native English speakers.
Authorship Signals (13%)
Detects indicators of AI-generated content or externally written resumes. Looks for unnaturally consistent tone, statistical language patterns typical of large language models, and stylistic uniformity that differs from natural human writing.
Metadata Analysis (5%)
Examines document properties - creator software, author field, creation and modification timestamps, and editing duration. A resume claiming 10 years of experience created in a resume builder 30 minutes before submission carries different weight than one maintained over time.

How Scoring Works

Scoring is deterministic given the same extracted facts. Once the LLM has extracted the structured facts from a resume, the score is computed entirely by code - no LLM judgment flows to the final number. Fact extraction runs at temperature 0 for stability, and we audit for cases where the extractor drifts.

Web Verification

When enabled, automated web searches check whether claimed employers and institutions exist online. A Bayesian statistical model adjusts the risk score based on verification results.

What is checked

What is NOT checked

Limitations

Web verification confirms that an entity (company, university) exists online. It does not confirm that the candidate actually worked at or attended that organization. Many legitimate small businesses have limited web presence - absence of results is explicitly treated as inconclusive, not as evidence of fabrication. All verification results should be confirmed through direct contact with employers and institutions.

Fairness and Bias

Non-Native English Speaker Adjustment

When the system detects that a resume is likely written by a non-native English speaker, scoring for Language Quality and Authorship Signals dimensions is reduced by up to 40%. Stylometric signals such as sentence structure and vocabulary diversity are unreliable indicators of fraud for non-native speakers.

Industry-Aware Scoring

Document metadata and web verification scoring is adjusted based on industry context. Technology companies (SaaS, fintech, cybersecurity) are expected to have strong web presence, so verification signals carry more weight. Companies in construction, agriculture, or manufacturing often have limited web presence, so scoring is reduced by 40% to avoid false positives.

Career Length Calibration

Template-like language is more forgivable for early-career candidates (1-5 years experience) and more concerning for senior professionals (10+ years). The system adjusts Language Quality scoring accordingly.

Four-Fifths Rule Monitoring

GetPruf includes an automated bias audit tool that analyzes scoring patterns across candidate segments. The audit applies the EEOC four-fifths rule to detect disparate impact: if a protected group's pass rate falls below 80% of the highest group's rate, the system flags the disparity for investigation. We run this audit regularly as data accumulates and adjust scoring weights when disparate impact is detected.

Bias Audit Results

Aggregated bias audit results are published as data accumulates. Early testing across synthetic datasets covering all major demographic segments showed no violations of the four-fifths rule. As real-world data grows, we will publish updated results on this page. If you have concerns about scoring fairness, contact us at support@getpruf.ai.

What GetPruf Is NOT

References

  1. Crosschq (2023). Reference Check Analytics: Patterns in Employment Verification.
  2. Loconte, A. et al. (2022). Resume Fraud Detection Methods: A Systematic Review. Journal of Business Ethics, 178(3), 601-618.
  3. Gartner (2025). Candidate Profile Fraud Forecast: 25% of profiles will be fabricated by 2028.
  4. Henle, C. et al. (2019). Resume Fraud Taxonomy. Personnel Psychology, 72(1). FAB 31%, EMB 72%, OMI 61%.
  5. Luo, Y. et al. (2018). ResumeNet: Cross-Section Consistency Analysis for Document Verification.
  6. EEOC (1978). Uniform Guidelines on Employee Selection Procedures. 29 CFR Part 1607.