SINGAPORE, Ju e 29, 2026 (GLOBE NEWSWIRE) — AIPOCH, i collabo atio with the Depa tme t of Pathology at Zho gsha Hospital, Fuda U ive sity, today u veiled MedSkillAudit, a p e-deployme t domai -specific audit f amewo k desig ed to ide tify scie tifically u eliable AI age t skills befo e they a e used i medical esea ch. The esea ch behi d it was published as a p ep i t o a Xiv (a Xiv:2604.20441) i Ap il 2026.
Medical esea ch age ts a e i c easi gly built f om modula skills that pe fo m tasks such as lite atu e sc ee i g, statistical a alysis, p otocol desig , a d ma usc ipt d afti g. Yet existi g quality-co t ol methods ofte fail to detect scie tific e o s, fab icated citatio s, flawed easo i g, o u safe outputs befo e these capabilities each esea che s.
MedSkillAudit i t oduces a two-laye “veto gate” eview p ocess. The fi st veto evaluates ope atio al stability, st uctu al co siste cy, esult dete mi ism, a d system secu ity, while the seco d assesses fou scie tific i teg ity dime sio s: scie tific i teg ity ( o fab icated citatio s, DOIs, sample sizes, o p-values), p actice bou da ies ( o di ect diag ostic co clusio s without p ope medical disclaime s), methodological baseli e ( o logical fallacies such as co flati g co elatio with causatio ), a d code usability ( o sy tax e o s o missi g co e depe de cies i ge e ated code). Skills that fail a y c itical equi eme t a e blocked f om deployme t.
Beyo d the two-laye veto gate, MedSkillAudit uses a two-stage evaluatio methodology: static evaluatio (desig quality, accou ti g fo 40%) a d dy amic evaluatio ( u time pe fo ma ce, accou ti g fo 60%).The f amewo k combi es a eview of the skill’s desig a d sou ce code with executio -based testi g i simulated esea ch sce a ios. Based o the fi al sco e, skills a e classified i to fou eadi ess levels a gi g f om “P oductio Ready”, “Limited Release” , “Beta O ly” to “Rejected”.
I a validatio study spa i g 75 skills ac oss five medical esea ch catego ies (e.g., evide ce I sight, p otocol desig , data a alysis, academic w iti g, a d othe ), 57.3% of skills fell below the Limited Release th eshold.The fi di gs highlight the u ge cy eed fo such gatekeepi g.
Mo e otably, the study also fou d that MedSkillAudit’s evaluatio s alig ed closely with expe t eviewe s a d delive ed co siste t esults ac oss diffe e t assessme ts.
“AI age ts a e becomi g pa t of the scie tific wo kflow, yet the e is still o equivale t of a quality-co t ol checkpoi t fo the skills they ely o ,” said Huimei Wa g, CEO at AIPOCH. “MedSkillAudit was developed to help esea che s ide tify scie tific, methodological, a d ethical isks befo e these capabilities a e deployed. We believe domai -specific auditi g f amewo ks will become a esse tial compleme t to t aditio al AI model evaluatio .”
A photo accompa yi g this a ou ceme t is available at https://www.globe ewswi e.com/NewsRoom/Attachme tNg/26e4fb94-a744-4125-b842-726178db5b86
