Question 1

Why does the project avoid a total match-score question?

Accepted Answer

The design keeps Jev on atomic judgments -- evidence gates, dimension scores and a routing choice -- and synthesizes the final match score, hard gates and bonuses in code, following the official guidance that Jev is best used for atomic questions combined by ordinary code.

Question 2

What changed from v1 to v3 of the criteria?

Accepted Answer

v2 disambiguated co-built versus owned metrics and coordination closure; v3 added evidence-over-self-description for AI use and candidate-attributable contribution for exceptional pursuits, dropping the trap resume's AI score from 0.95 to 0.49.

Question 3

What are the ten questions in the one request?

Accepted Answer

Five Noul gates for education floor, written English evidence, heavy AI product use, efficiency tools in work and exceptional pursuit; four Score questions on 0-3 scales for user community work depth, data and metrics ability, cross-team coordination delivery and signal-to-insight; and one Choice that buckets the dominant background for routing to a recruiter. Instructions and criteria are written in English.

Question 4

What did the calibration profile look like on the positive sample?

Accepted Answer

Fact-retrieval questions such as degree and certificates came back at 0.97-0.99 confidence while the subjective exceptional-pursuit question returned 0.89-0.90 for the same evidence. The coordination question did not round up to 3 but sat on the fence at {2: 0.33, 3: 0.67} with 0.67 confidence, which the README reads as the model knowing it is a borderline case.

Question 5

What is the cross-language risk, and how is it handled?

Accepted Answer

The same resume scored 0.67 confidence in Chinese and 0.75 in English, so a 0.7 routing threshold would send the Chinese version to a human and auto-advance the English one. The fix is to fix one pipeline language and calibrate the thresholds in that language. Chinese also costs about 10% more tokens and 1.9 times the latency.

Question 6

What does a resume cost, and what are the compliance guardrails?

Accepted Answer

About $0.00016 per resume, at 3,901 tokens and 197 ms for Chinese against 3,525 tokens and 106 ms for English. Before sending, strip name, photo, gender, age and marital status; every Noul asks whether the resume states X, with background checks as a separate step; model output ranks only and never auto-rejects, and rejections are sampled for human review because confidence defenses fail on out-of-distribution input.

Question 7

What compliance red lines does the project set for itself?

Accepted Answer

Strip PII before sending: name, photo, gender, age and marital status. Every Noul tests whether the resume states something, not whether it is true, so background checks stay a separate step. Model output ranks only and never auto-rejects, because recruiting is a high-risk AI Act scenario, and rejections are sampled for human review since confidence defenses fail on out-of-distribution input. There is no license file, so reuse terms are unclear.

Jev Resume Screening

Highlights

Questions about this entry

Related terms

More Jev resources like this

From the Jev guides

From the Jev community