FreeCVAudit accuracy benchmark

Interpretation and limitations

Each sample has one controlled target alteration or no deliberately planted issue. The detection measurement tests whether the corresponding rule fires; the visible measurement tests whether it is among the three or four highest-priority findings shown to the user. This is not comprehensive human annotation, and the measured ratio is not general CV-audit accuracy, sensitivity, specificity or an ATS pass rate.

The free tool analyses extracted text, not PDF visual geometry. It cannot verify visual reading order, hidden text, tables reconstructed incorrectly, or a particular recruitment platform's interpretation. English and French section patterns are better represented in the current heuristics than Arabic or Spanish; a multilingual document can be misclassified. Internal diagnostic scores are capped and do not represent calibrated probabilities of recruitment success.

For scanned documents, the free audit reports the absence of text. Following payment, the separate Gemini-powered workflow can transcribe the scan before generating a revised CV. That capability should not be confused with the free audit's scan detection, and its accuracy requires a separate verified evaluation with transcription-level ground truth. No OCR accuracy percentage is claimed here.

Benchmark operator: FreeCVAudit.com · synthetic input set · version 2026-10-29. Report a discrepancy or request a re-test.