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-2026-2458
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-10T18:02:18.704128Z
status: published
lens: g_space
sector: health
headline: Coexistence of AI Nurse and AI-Simulated Patient in Nursing Education Using a Bidirectional Fine-Tuning Pipeline: Development and Content Analysis Study
dek: Background Simulated patient systems are essential in nursing education but face actor scarcity, instructor burden, cost, and disrupted continuity where in-person simulation cannot be sustained. Existing large language model (LLM)-based systems address these issues but treat the dialogic partner-the AI nurse-as a static prompt or unspecialized model, compromising data collection quality and pedagogical applicability. Objective We report (1) the construction of 2 specialized fine-tuned models, NURSE-ONE (AI nurse…
gain_title: In an exploratory evaluation, ten nursing faculty rated a fine-tuned AI-simulated patient at 90.3%, the pairing of AI patient and AI nurse at 84.8%, and the AI nurse at 69.0%, with all three modes responding within 3 seconds on one consumer graphics card.
problem_title: (none)
trace_subject: (none)
gain_reading: In an exploratory evaluation, ten nursing faculty rated a fine-tuned AI-simulated patient at 90.3%, the pairing of AI patient and AI nurse at 84.8%, and the AI nurse at 69.0%, with all three modes responding within 3 seconds on one consumer graphics card.
gain_evidence: Third-party evaluation rated the AI-simulated patient, AI nurse, and dyadic interaction at 90.3%, 69.0%, and 84.8%, respectively | All 3 modes met the prespecified 3-second latency threshold
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers built two specialized AI models for nursing education: NURSE-ONE, an AI nurse, and PATIENT-ONE, an AI-simulated patient. Ten nursing faculty tested them in a working web interface and rated the AI patient at 90.3%, the AI nurse at 69.0% and the interaction between the two at 84.8%. Role maintenance was rated at the ceiling (mean 5, SD 0).

The models were fine-tuned on a single consumer-grade Nvidia RTX 4090 with 24 GB of memory, starting from the Gemma-based gemma4-E4B-it model. The system supports three operating modes, including offline use for disaster preparedness, and all three met a 3-second response-time threshold.

This is an early development study. Faculty ratings came from ten evaluators, and the authors describe their findings on how the two models differ as a preliminary pattern awaiting replication. No student learning outcomes have been measured yet.
limitation: The faculty evaluation was exploratory with ten evaluators, the reported asymmetry between the two models is a preliminary pattern pending replication, and learner outcomes have not been studied.
tag: Evidence-backed gain
key_points: The training base was 251 scenarios drawn from 29 institutionally authored nursing care plans and 20 items from the 115th Japanese National Nursing Examination. | The AI nurse was fine-tuned with 245, 735 and 1225 cases; the patient model was adopted at 735 cases, which gave the lowest response-length variability (SD 24.5) and the highest information-providing rate (47.9%). | Two independent nursing experts coded 196 blinded utterances (κ=0.770), and two AI raters largely agreed with them (κ=0.742). | Ten nursing faculty scored the system on a 21-item rubric with an item-level content validity index of 1.00. | The AI nurse scored lowest of the three components at 69.0%, below the AI patient at 90.3%. | All three operating modes met the 3-second latency threshold on consumer hardware and can run offline.
rundown: The five-step pipeline first built a scenario base, then used a provisional dialogue system (AutoGen v18, maxmsg=25) to collect training dialogues. NURSE-ONE was fine-tuned across three training volumes, and PATIENT-ONE was then fine-tuned on dialogues generated by NURSE-ONE.

Content analysis compared the conditions using the Berelson method. The patient-side comparison showed no significant category shift (P=.30), but the 735-case condition had the steadiest response lengths. The authors say the nurse adapter mainly changed questioning style while the patient adapter mainly stabilised responses, which they call functional asymmetry.

The faculty evaluation covered three coexistence modes in a web interface. The authors say it supports exploratory educational applicability and provides a foundation for future learner outcome studies, which have not yet been run.
sources:
- peer_reviewed | JMIR Biomedical Engineering | https://doi.org/10.2196/100778 | 2026-11-06
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

ac53b309cc44c093d07fdf8c45bbef02a2437e4e210cb82344c027cd18282671

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000