Litmus

Technical improvements for more precise review

The same word can appear in both a problematic disclosure and an ordinary sentence. Litmus combines explicit rules, AI context analysis, and exception handling, with results that can be traced back to the original passage.

Improvement	Approach	What reviewers can check
Separate rules from AI interpretation	Calculate length, blank answers, and repetition with rules; use AI where meaning matters	Which criteria were applied to each check
Require two plagiarism conditions	Check similarity together with a minimum length of consecutive matching text	Common wording and the passages that actually overlap
Apply contextual exceptions	Distinguish profanity in use, organizations in work history, and anonymized names	Why an otherwise normal expression was flagged
Structure the evidence	Connect applicant, question, source extract, check item, and reason	Each flag alongside its original text
Manage institution settings and versions	Record guidance, rules, and model versions separately for each institution	The criteria used for that recruitment review

Check more than similarity for suspected plagiarism

Applications often share job titles and common ways of describing experience. Plagiarism checks were refined to require both similarity and consecutive matching text. This is designed to avoid flagging a passage just because it uses familiar words, while helping reviewers inspect the actual overlap.

Flagged results link to the question and source evidence. Reviewers assess the wording in context against institutional guidance before making a decision.

Reduce flags on ordinary expressions

• Profanity: distinguish a matching string inside an ordinary Korean phrase from an actual insult.
• Institution names: account for previous employers, anonymized names such as “○○,” and generic references to the hiring organization.
• Blind-recruitment checks: distinguish birthplace from current residence, or general military service from a specific role, according to the institution’s guidance.

These exceptions are designed to reduce unnecessary rechecking. Each institution configures the exceptions and interpretation it needs.

Rules and contextual review

Check	Processing and review approach
Length, blank answers, repetition	Explicit counts and repeat-pattern checks
Blind-recruitment disclosures	Institution-defined criteria applied with context
Incorrect institution names	Distinguish work history or anonymized wording from a potential error
Suspected plagiarism	Review similarity together with consecutive matching text
Profanity	Distinguish matching strings from actual usage in context

Whether birthplace and current residence should be treated alike depends on institutional guidance. Military service and relatives’ employment are also policy choices, not universal defaults.

Evidence and operating criteria for reviewers

Results connect the applicant, question, original passage, check item, and reason. Reviewers compare these with the guidance and can revisit the same evidence in later questions or appeals.

Criteria, rules, and model versions are recorded, with settings separated by institution. Applicant data is used for the relevant recruitment review, and retention and deletion follow institutional policy. Reviewers decide eligibility and recruitment outcomes.

Manage analysis cost alongside quality

Prompt caching reuses recurring guidance, while compatible checks are combined into one model call to reduce repeated requests. In an implementation using these optimizations, AI call costs fell by approximately 40%. This figure measures AI call costs, not total costs including human review.

Candidate model changes are compared on the same applications for false flags, missed issues, and output structure before adoption.

A pilot uses actual application samples and institutional guidance to check whether the reasons shown are appropriate, whether issues were missed, and how much correction reviewers need. The criteria are refined using those findings.