Benchmark
In this benchmark, CVault won 74% of contested fields and 8 resumes, with 2 ties. This page keeps the result visible while avoiding old oversized benchmark styling.
Historical internal comparison. This ten-template-resume run was not independently audited and is not a claim about every parser or production corpus.
CVault win rate on contested field comparisons.
2 resumes tied and 0 lost in this benchmark set.
CVault field wins compared with Affinda field wins.
| Field | CVault wins | Affinda wins | Ties |
|---|---|---|---|
| Responsibilities | 10 | 0 | 0 |
| Education | 6 | 0 | 3 |
| Phone | 5 | 0 | 4 |
| Certifications | 5 | 1 | 3 |
| Job titles | 3 | 0 | 6 |
| 2 | 0 | 7 | |
| Summary / Profile | 1 | 0 | 8 |
| Skills quantity | 2 | 5 | 2 |
| 0 | 1 | 9 |
Every resume produces clean JSON arrays of individual bullet points. Affinda returns unstructured text blobs with \n separators, requiring downstream re-parsing. This is the single biggest structural advantage for ATS integration, search indexing, and evidence-backed candidate matching.
CVault captured all education entries on every resume, including edge cases: General Studies at a community college, a 3-entry section on a 3-page executive resume, and 5 entries across 3 countries. Affinda missed entries on multiple resumes and misclassified certifications as education.
Affinda dropped phone numbers on 5 of 10 resumes. CVault captured every phone number present on every resume.
CVault correctly separates certifications from education. Affinda folds certifications into education entries with a "Course/Certificate" level tag, conflating two distinct candidate data categories.
CVault preserves full job titles including geographic scope ("HEAD OF HR, INDIA"). Affinda splits geo-scoped titles incorrectly, assigns the geographic region as the employer, adds trailing commas, or drops titles entirely.
On one resume where no email was present, CVault hallucinated one. Affinda correctly returned nothing. A real edge case to fix.
Affinda generates more skills by inferring aggressively from job descriptions (e.g., 138 skills on a 1-page resume). CVault extracts skills closer to what is explicitly listed. More inferred skills means more noise alongside more signal — a design tradeoff, not an accuracy gap.
Both parsers produce identical errors on OCR-damaged PDFs ("SACETY COORDINATOR", "DADA/DACA SCIENTIST"). This is an upstream text extraction issue, not parser logic.
When resumes omit company names entirely, neither parser can invent them.
Both miss these non-standard fields.
