extract: 2025-11-01-jmir-knowledge-practice-gap-39-benchmarks-systematic-review

Pentagon-Agent: Epimetheus <3D35839A-7722-4740-B93D-51157F7D5E70>

domains/health/medical LLM benchmark performance does not translate to clinical impact because physicians with and without AI access achieve similar diagnostic accuracy in randomized trials.md

@@ -35,6 +35,12 @@ OpenEvidence's medRxiv preprint (November 2025) showed 24% accuracy for relevant
 
 ARISE report identifies specific failure modes: real-world performance 'breaks down when systems must manage uncertainty, incomplete information, or multi-step workflows.' This provides mechanistic detail for why benchmark performance doesn't translate — benchmarks test pattern recognition on complete data while clinical care requires uncertainty management.
 
+### Additional Evidence (extend)
+*Source: [[2025-11-01-jmir-knowledge-practice-gap-39-benchmarks-systematic-review]] | Added: 2026-03-24*
+
+JMIR systematic review of 761 studies provides methodological foundation: 95% of clinical LLM evaluation uses medical exam questions rather than real patient data, with only 5% assessing performance on actual patient care. Traditional benchmarks show saturation at 84-90% USMLE accuracy, but conversational frameworks reveal 19.3pp accuracy drop (82% → 62.7%) when moving from case vignettes to multi-turn dialogues. Review concludes: 'substantial disconnects from clinical reality and foundational gaps in construct validity, data integrity, and safety coverage.' This establishes that the Oxford/Nature Medicine RCT deployment gap (94.9% → 34.5%) is part of a systematic field-wide pattern, not an isolated finding.
+
+
 
 
 Relevant Notes:

inbox/archive/general/2025-10-15-cell-reports-medicine-llm-pharmacist-copilot-medication-safety.md

@@ -0,0 +1,57 @@
+---
+type: source
+title: "Cell Reports Medicine 2025: Pharmacist + LLM Co-pilot Outperforms Pharmacist Alone by 1.5x for Serious Medication Errors"
+author: "Multiple authors (Cell Reports Medicine, cross-institutional)"
+url: https://pmc.ncbi.nlm.nih.gov/articles/PMC12629785/
+date: 2025-10-15
+domain: health
+secondary_domains: [ai-alignment]
+format: research-paper
+status: processed
+priority: medium
+tags: [clinical-ai-safety, centaur-model, medication-safety, llm-copilot, pharmacist, clinical-decision-support, rag, belief-5-counter-evidence]
+---
+
+## Content
+
+Published in *Cell Reports Medicine*, October 2025 (doi: 10.1016/j.xcrm.2025.00396-9). Prospective, cross-over study. Published in PMC as PMC12629785.
+
+**Study design:**
+- 91 error scenarios based on 40 clinical vignettes across **16 medical and surgical specialties**
+- LLM-based clinical decision support system (CDSS) using retrieval-augmented generation (RAG) framework
+- Three arms: (1) LLM-based CDSS alone, (2) Pharmacist + LLM co-pilot, (3) Pharmacist alone
+- Outcome: accuracy in identifying medication safety errors
+
+**Key findings:**
+- **Pharmacist + LLM co-pilot:** 61% accuracy (precision 0.57, recall 0.61, F1 0.59)
+- **Serious harm errors:** Co-pilot mode increased accuracy by **1.5-fold over pharmacist alone**
+- Conclusion: "Effective LLM integration for complex tasks like medication chart reviews can enhance healthcare professional performance, improving patient safety"
+
+**Implementation note:** This used a RAG architecture (retrieval-augmented generation), meaning the LLM retrieved drug information from a curated database rather than relying solely on parametric memory — reducing hallucination risk.
+
+## Agent Notes
+
+**Why this matters:** This is the clearest counter-evidence to Belief 5's pessimistic reading in the KB. Where NOHARM shows 22% severe error rates and the Oxford RCT shows zero improvement over controls, this study shows a POSITIVE centaur outcome: pharmacist + LLM outperforms pharmacist alone by 1.5x on the outcomes that matter most (serious harm errors). This is the centaur model working as intended.
+
+**What surprised me:** The 1.5x improvement on serious harm specifically — not just average accuracy. This means the LLM helps most where the stakes are highest. That's the ideal safety profile: catching the worst errors. The RAG architecture may be key — this isn't a general chat LLM but a structured decision support tool with constrained information retrieval.
+
+**What I expected but didn't find:** A clear statement of failure conditions. When does the co-pilot model FAIL to improve? The 61% accuracy ceiling suggests the co-pilot mode also misses ~39% of errors. The study doesn't clearly delineate what the LLM adds vs. what it misses.
+
+**KB connections:**
+- Counter-evidence to Sessions 8-11 clinical AI safety concern: the centaur model CAN work in specific conditions (RAG architecture, domain-expert+LLM combination, structured safety task)
+- The centaur design requires domain expert + LLM — this is specifically a pharmacist co-pilot, not a physician being replaced
+- Connects to NOHARM: NOHARM found 76.6% of severe errors are omissions. If the pharmacist+LLM catches errors the pharmacist alone misses, the omission-detection mechanism is real — but requires the pharmacist to be present and engaged (not automation bias mode)
+- The RAG architecture is important: this isn't vulnerable to the misinformation propagation failure mode (Lancet DH 2026) the way a general LLM is, because it retrieves from a curated database
+- Connects to the distinction between "clinical reasoning AI" (OE) and "structured CDSS with RAG" (this study) — these are different products with different safety profiles
+
+**Extraction hints:**
+- Primary claim: "LLM-based clinical decision support in co-pilot mode with a domain expert improves serious medication harm detection by 1.5x vs. pharmacist alone — evidence that centaur design works for structured safety tasks using RAG architecture"
+- The constraint is important: centaur works when (a) the expert is engaged (not automation bias mode), (b) the LLM uses RAG (not parametric memory), (c) the task is structured (medication safety, 16 specialties)
+- This limits the claim — it does NOT say "clinical AI is safe in general" — it says "LLM + expert in a structured RAG setting improves safety for a defined task"
+
+**Context:** Cell Reports Medicine is a high-tier Cell Press journal for clinical translational research. Prospective cross-over design with clear comparison arms. 16 specialties gives the finding breadth across clinical contexts.
+
+## Curator Notes
+PRIMARY CONNECTION: Belief 5 counter-evidence — centaur model works under specific conditions
+WHY ARCHIVED: Best positive clinical AI safety evidence found across 12 sessions; establishes the conditions under which centaur design improves outcomes
+EXTRACTION HINT: Extract with explicit scope constraint: centaur + RAG + structured safety task = works; general CDSS + automation bias mode = doesn't work per other evidence

inbox/queue/.extraction-debug/2025-10-15-cell-reports-medicine-llm-pharmacist-copilot-medication-safety.json

@@ -0,0 +1,24 @@
+{
+  "rejected_claims": [
+    {
+      "filename": "llm-copilot-with-rag-architecture-improves-pharmacist-medication-error-detection-by-1-5x-for-serious-harm-cases.md",
+      "issues": [
+        "missing_attribution_extractor"
+      ]
+    }
+  ],
+  "validation_stats": {
+    "total": 1,
+    "kept": 0,
+    "fixed": 1,
+    "rejected": 1,
+    "fixes_applied": [
+      "llm-copilot-with-rag-architecture-improves-pharmacist-medication-error-detection-by-1-5x-for-serious-harm-cases.md:set_created:2026-03-24"
+    ],
+    "rejections": [
+      "llm-copilot-with-rag-architecture-improves-pharmacist-medication-error-detection-by-1-5x-for-serious-harm-cases.md:missing_attribution_extractor"
+    ]
+  },
+  "model": "anthropic/claude-sonnet-4.5",
+  "date": "2026-03-24"
+}

inbox/queue/.extraction-debug/2025-11-01-jmir-knowledge-practice-gap-39-benchmarks-systematic-review.json

@@ -0,0 +1,32 @@
+{
+  "rejected_claims": [
+    {
+      "filename": "clinical-llm-evaluation-uses-medical-exam-questions-not-real-patient-data-creating-systematic-benchmark-validity-gap.md",
+      "issues": [
+        "missing_attribution_extractor"
+      ]
+    },
+    {
+      "filename": "conversational-clinical-ai-shows-19-point-accuracy-drop-versus-single-turn-questions-revealing-interaction-complexity-gap.md",
+      "issues": [
+        "missing_attribution_extractor"
+      ]
+    }
+  ],
+  "validation_stats": {
+    "total": 2,
+    "kept": 0,
+    "fixed": 2,
+    "rejected": 2,
+    "fixes_applied": [
+      "clinical-llm-evaluation-uses-medical-exam-questions-not-real-patient-data-creating-systematic-benchmark-validity-gap.md:set_created:2026-03-24",
+      "conversational-clinical-ai-shows-19-point-accuracy-drop-versus-single-turn-questions-revealing-interaction-complexity-gap.md:set_created:2026-03-24"
+    ],
+    "rejections": [
+      "clinical-llm-evaluation-uses-medical-exam-questions-not-real-patient-data-creating-systematic-benchmark-validity-gap.md:missing_attribution_extractor",
+      "conversational-clinical-ai-shows-19-point-accuracy-drop-versus-single-turn-questions-revealing-interaction-complexity-gap.md:missing_attribution_extractor"
+    ]
+  },
+  "model": "anthropic/claude-sonnet-4.5",
+  "date": "2026-03-24"
+}

inbox/queue/2025-10-15-cell-reports-medicine-llm-pharmacist-copilot-medication-safety.md

@@ -7,9 +7,13 @@ date: 2025-10-15
 domain: health
 secondary_domains: [ai-alignment]
 format: research-paper
-status: unprocessed
+status: null-result
 priority: medium
 tags: [clinical-ai-safety, centaur-model, medication-safety, llm-copilot, pharmacist, clinical-decision-support, rag, belief-5-counter-evidence]
+processed_by: vida
+processed_date: 2026-03-24
+extraction_model: "anthropic/claude-sonnet-4.5"
+extraction_notes: "LLM returned 1 claims, 1 rejected by validator"
 ---
 
 ## Content
@@ -55,3 +59,11 @@ Published in *Cell Reports Medicine*, October 2025 (doi: 10.1016/j.xcrm.2025.003
 PRIMARY CONNECTION: Belief 5 counter-evidence — centaur model works under specific conditions
 WHY ARCHIVED: Best positive clinical AI safety evidence found across 12 sessions; establishes the conditions under which centaur design improves outcomes
 EXTRACTION HINT: Extract with explicit scope constraint: centaur + RAG + structured safety task = works; general CDSS + automation bias mode = doesn't work per other evidence
+
+
+## Key Facts
+- Cell Reports Medicine published prospective cross-over study in October 2025 (doi: 10.1016/j.xcrm.2025.00396-9, PMC12629785)
+- Study tested 91 error scenarios based on 40 clinical vignettes across 16 medical and surgical specialties
+- Pharmacist + LLM co-pilot achieved 61% accuracy (precision 0.57, recall 0.61, F1 0.59)
+- Co-pilot mode increased accuracy by 1.5-fold over pharmacist alone for serious harm errors specifically
+- LLM-based CDSS used retrieval-augmented generation (RAG) framework with curated drug database

inbox/queue/2025-11-01-jmir-knowledge-practice-gap-39-benchmarks-systematic-review.md

@@ -7,9 +7,13 @@ date: 2025-11-01
 domain: health
 secondary_domains: [ai-alignment]
 format: research-paper
-status: unprocessed
+status: enrichment
 priority: medium
 tags: [clinical-ai-safety, benchmark-performance-gap, llm-evaluation, knowledge-practice-gap, real-world-deployment, belief-5, systematic-review]
+processed_by: vida
+processed_date: 2026-03-24
+enrichments_applied: ["medical LLM benchmark performance does not translate to clinical impact because physicians with and without AI access achieve similar diagnostic accuracy in randomized trials.md"]
+extraction_model: "anthropic/claude-sonnet-4.5"
 ---
 
 ## Content
@@ -53,3 +57,14 @@ Published in *Journal of Medical Internet Research* (JMIR), 2025, Vol. 2025, e84
 PRIMARY CONNECTION: Belief 5 — clinical AI safety evaluation methodology gap
 WHY ARCHIVED: Provides systematic evidence that the KB's reliance on benchmark performance data (e.g., "OE scores 100% on USMLE") is epistemically weak — and establishes that the Oxford RCT deployment gap finding is part of a systematic pattern
 EXTRACTION HINT: Extract the 5%/95% finding as a standalone methodological claim about the clinical AI evaluation field; pair with Oxford Nature Medicine RCT as empirical confirmation
+
+
+## Key Facts
+- JMIR systematic review analyzed 761 LLM evaluation studies across 39 benchmarks
+- Only 5% of 761 studies assessed performance on real patient care data
+- 95% of studies relied on medical examination questions (USMLE-style) or case vignettes
+- Leading models achieve 84-90% accuracy on USMLE benchmarks
+- Diagnostic accuracy drops from 82% on case vignettes to 62.7% on multi-turn dialogues (19.3pp decrease)
+- npj Digital Medicine study: six LLMs averaged 57.2% total score, 54.7% safety score, 62.3% effectiveness
+- 13.3% performance drop in high-risk scenarios versus average scenarios (npj Digital Medicine)
+- LLMs show markedly lower performance on script concordance testing than on multiple-choice benchmarks