CVoria publishes privacy model benchmark whitepaper
Our whitepaper shows how well CVoria's own privacy model replaces personal details in CVs before AI analysis, and where it still falls short.
23 of 757
Core personal details CVoria's privacy model left readable in 200 new fictional CVs.
108 of 757
The best open tool tested, GLiNER multi PII, left more than four times as many readable.
7 and 17
Gemini 3.8 Flash and GPT-5.6 Luna, given our redaction rules, did better, but they have to receive the full CV.
200 CVs
Written by hand for this benchmark by two AI models, labelled twice and locked before any system ran.
The headline result
Before CVoria sends CV text to an AI provider, its own privacy model replaces the personal details it finds with placeholders. On 200 new fictional CVs it left 23 of 757 core personal details readable: names, email addresses, phone numbers, street addresses, personal links, dates of birth and ID numbers.
GLiNER multi PII left 108, LiquidAI LFM2.5 PII 199 and Microsoft Presidio 299. Our model left no phone number and no ID number readable; most of its misses were names in unusual forms, such as children's first names and nicknames.
How the benchmark worked
Claude Opus 5.5 and GPT-6.1 Sol each wrote 100 fictional CVs by hand, in Swedish, English and 15 other languages, with contact details that cannot reach anyone. Each model then labelled the other's CVs without seeing the first answer key.
The test set and the scoring were locked with checksums before any system ran. The measure is how many personal details can still be read after each system's output is blanked out of the CV.
Where the hosted models did better
Given our written redaction rules, Gemini 3.8 Flash left 7 core details readable and GPT-5.6 Luna 17, and both were more precise than our model. To do that they have to receive the full CV, personal details included. CVoria's model runs in our own AWS account in Stockholm, so the details it finds are replaced before the text reaches an AI provider.
Our model also sometimes removes dates that are not birth dates, which the next model update will fix.
What is in the public package
The package has all 200 CVs with their answer keys, every system's output, both labellers' keys and the decisions between them, the writing instructions and the scoring scripts. One command recalculates every table in the whitepaper.
The CVs are fictional and written by AI models, so results on real CVs can differ. The whitepaper lists the limits in full.