The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2864
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T18:02:25.600086Z
status: published
lens: g_space
sector: health
headline: Generative AI in the Taiwan medical laboratory scientist examination: Performance and educational utility
dek: The rapid evolution of generative artificial intelligence (GAI), particularly large language models (LLMs), offers transformative potential for health professions education. However, their performance in specialized clinical domains, such as medical laboratory science as well as educational utility, remain insufficiently characterized. This study evaluated multiple LLMs on the Taiwanese Medical Laboratory Scientist Examination (TMLSE) and investigated the effectiveness of AI-generated assessments in undergraduat…
gain_title: In a study testing four AI models on six Taiwanese Medical Laboratory Scientist Examination administrations, GPT-5.2 scored 93.5% and Gemini 3 Pro 92.1%, and 30 students rated AI-generated items significantly higher than real exam questions for overall satisfaction (p=0.0089).
problem_title: (none)
trace_subject: (none)
gain_reading: In a study testing four AI models on six Taiwanese Medical Laboratory Scientist Examination administrations, GPT-5.2 scored 93.5% and Gemini 3 Pro 92.1%, and 30 students rated AI-generated items significantly higher than real exam questions for overall satisfaction (p=0.0089).
gain_evidence: GPT-5.2 (93.5%) and Gemini 3 Pro (92.1%) significantly outperformed GPT-4o (85.8%) and GPT-3.5 (67.1%) | UGY students rated AI-generated items significantly higher than authentic TMLSE questions did in terms of overall satisfaction (p= 0.0089)
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers tested four AI models (GPT-3.5, GPT-4o, GPT-5.2 and Gemini 3 Pro) on six administrations of the Taiwanese Medical Laboratory Scientist Examination from 2021 to 2023. All four cleared the 60% passing threshold, with GPT-5.2 scoring 93.5%, Gemini 3 Pro 92.1%, GPT-4o 85.8% and GPT-3.5 67.1%.

The team also had experts validate 180 AI-generated exam items and gave them to 30 undergraduate-year medical laboratory students. The students rated the AI items higher than real exam questions for overall satisfaction and for usefulness in exam review, and scores on the AI items tracked clinical internship outcomes.

Image interpretation remained the weak point: GPT-5.2 reached 73.3% on image-based items. Informational gaps accounted for more than 95% of wrong answers across all models.
limitation: The educational part involved only 30 undergraduate-year students, and the correlation with internship outcomes was modest (r=0.3662, p=0.0465); image-based questions remained a weakness for the models.
tag: Evidence-backed gain
key_points: Four models were tested on six exam administrations from 2021 to 2023, across 10 disciplines plus image-based items. | All four models passed the 60% threshold; GPT-3.5 scored lowest at 67.1%. | GPT-5.2 (93.5%) and Gemini 3 Pro (92.1%) significantly outperformed GPT-4o (85.8%) and GPT-3.5 (67.1%). | Informational deficiencies caused more than 95% of incorrect answers; the rest were logical or statistical errors. | Students rated AI-generated items higher than real exam questions for satisfaction (p=0.0089) and exam-review utility (p=0.0405). | Scores on AI-generated assessments correlated with clinical internship outcomes (Spearman's r=0.3662, p=0.0465); traditional exam scores showed no significant association.
rundown: The evaluation covered six administrations of Taiwan's national medical laboratory scientist exam. Wrong answers were sorted into informational, logical or statistical errors, and informational gaps dominated at more than 95% across all models.

For the education part, 180 AI-generated items were validated by experts and given to 30 undergraduate-year students. They rated these items significantly higher than authentic exam questions for overall satisfaction (p=0.0089) and usefulness for review (p=0.0405).

Performance on the AI-generated assessments was positively correlated with clinical internship outcomes (r=0.3662, p=0.0465), while traditional exam scores showed no significant association. The authors conclude that image interpretation remains a limitation for the models.
sources:
- peer_reviewed | Journal of the Chinese Medical Association | https://doi.org/10.1097/jcma.0000000000001445 | 2026-11-12
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

94d614c2f6fd1dce5b322a1551d86ea93ca611d9a2384a9e5aa06796baf1c7a1

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000