DeepLens Diagnosis Agent: Agentic Workflow Design Lets a Small Reasoning Model Compete with Frontier LLMs

# DeepLens Diagnosis Agent: Agentic Workflow Design Lets a Small Reasoning Model Compete with Frontier LLMs
Source: [https://arxiv.org/html/2607.22555](https://arxiv.org/html/2607.22555)
\\KOMAoptions

headsepline=false,footsepline=false

\\conference

Veysel KocamanMuhammed Ali NaqviYigit GulDavid TalbyJohn Snow Labs

###### Abstract

Medical diagnosis is a multi\-stage process: extract facts, consult knowledge, generate a differential analysis, and select the best diagnosis with explanations\. Frontier LLMs are strong generalists, but single\-shot prompting often yields brittle diagnostic reasoning\. We present the DeepLens Diagnosis Agent, a five\-stage*harnessing*pipeline—the practice of combining model capabilities with disciplined process constraints—centered on a small medical reasoning model \(JSL Medical Small 7B v2\)\[jsl\_medllm\_docs\]and retrieval\-augmented generation \(RAG\)\[lewis2020rag\]\. The pipeline enforces structured clinical extraction, disciplined retrieval, constrained candidate generation, explicit evidence triangulation, and an auditable final decision\. On a held\-out 915\-case diagnostic benchmark \(DiagnosisArena\)\[zhu2025diagnosisarena\], the agent achieved60\.14%top\-1 diagnostic accuracy, the highest among small and medium\-sized models\. The same JSL Medical Small 7B v2 model without the agent workflow achieved23\.99%, a\+36\-point gainattributable to workflow design alone, despite the base model achieving 88\.2% on standard medical benchmarks—demonstrating that diagnostic reasoning under uncertainty requires more than medical knowledge recall\. This accuracy gain is achieved at$0\.0072 per case\(24K tokens per case on A100 infrastructure\) with 24\-second latency—35–45% cheaper than frontier cloud models \(Claude Sonnet 4\.5: $0\.0110, Gemini 3\.1 Pro: $0\.0128\) while outperforming them by \+9\.70pp and \+9\.17pp respectively\. Importantly, we demonstrate that harnessing and agentic diagnosis can improve frontier model failures, suggesting that workflow discipline is a complementary lever for all model scales\. For diagnostic reasoning, workflow constraints and verification gates can outweigh parameter count or API costs\. Beyond aggregate accuracy, the pipeline produces structured intermediate artifacts that make each stage inspectable and facilitate error localization during failure analysis\. These properties are essential for high\-stakes deployment settings, where traceability, reproducibility, and reviewer\-auditable evidence are necessary alongside strong benchmark performance\.

###### keywords:

medical diagnosis\\sepagentic workflow\\sepretrieval\-augmented generation\\sepdifferential diagnosis\\sepclinical reasoning

## 1Introduction

Medical diagnosis is fundamentally a workflow problem\. Clinicians do not produce diagnoses in one step: they structure patient facts, consult knowledge, generate and refine a differential diagnosis, and finally commit to a decision\.

In contrast, many LLM baselines attempt implicit multi\-step reasoning within a single completion\. This tends to blur*extraction*\(what is stated\) and*inference*\(what is implied\), increasing error propagation\.

DiagnosisArena emphasizes this difficulty: long clinical narratives require precise fact grounding \(positives, negatives, timeline, pathology\) and multi\-step inference to reach the correct diagnosis\[zhu2025diagnosisarena\]\. More broadly, the field is trending toward structured evaluation and disciplined medical reasoning workflows \(e\.g\., medical reasoning foundation models\[octomed2025\]and expert review protocols\)\[kocaman2025clever\]\.

This paper advances the following thesis: a small reasoning model embedded in a carefully engineered workflow can outperform larger models when the system design is explicitly constrained\. We call this approach*harnessing*—the practice of combining model capabilities with disciplined process constraints to achieve reliable, auditable reasoning at scale\.

The motivation is practical as well as scientific\. In real clinical settings, diagnostic assistants are expected to be traceable, stable across repeated runs, and easy to audit when outputs are disputed\. Single\-pass free\-form generation tends to optimize fluency, whereas clinical deployment requires explicit control over evidence grounding, error localization, and decision criteria\. The workflow perspective adopted in this paper addresses that requirement by making each stage inspectable and by constraining how intermediate outputs can influence the final diagnosis\.

Accordingly, this paper focuses on system\-level design choices rather than model scaling alone \(Figure[1](https://arxiv.org/html/2607.22555#S2.F1)provides a high\-level overview of the proposed pipeline\)\. We show how decomposition into extraction, supportive retrieval, constrained candidate generation, and quote\-anchored triangulation changes diagnostic behavior in measurable ways\. We then evaluate this design on DiagnosisArena and discuss why the observed gains are consistent with workflow\-level error reduction, not just stronger parametric recall\.

From an engineering perspective, this design can be interpreted as a set of explicit reliability constraints applied to each reasoning stage\. Instead of optimizing for a single fluent output, the workflow optimizes for traceable intermediate artifacts that can be validated and audited\. This shift is central to clinical decision support, where interpretability and failure localization are often as important as benchmark\-level accuracy\.

## 2Background and Related Work

Recent progress in medical LLMs shows strong gains on benchmarked QA, but also highlights a persistent gap between benchmark accuracy and reliable clinical reasoning under uncertainty\. Early evidence that general\-purpose LLMs encode clinically useful knowledge came from MultiMedQA/Med\-PaLM style evaluations\[singhal2023medpalm\], and later work improved specialist\-grade performance via better medical alignment and prompting strategies\[tu2025expertlevel\],\[nori2023medprompt\]\. In parallel, benchmark design evolved toward harder diagnostic settings that require structured synthesis over long clinical narratives instead of short fact recall\[zhu2025diagnosisarena\]\.

For this paper, the key background assumption is that diagnosis quality depends not only on model knowledge, but also on reasoning\-process control\. Chain\-of\-thought prompting showed that explicit intermediate reasoning can improve task performance\[wei2022cot\],\[wang2022selfconsistency\], while ReAct\-style formulations demonstrated the value of combining reasoning with external evidence access\[yao2022react\]\. In medical tasks, these findings imply that architecture\-level constraints \(fact extraction, grounded comparison, and explicit tie\-breaks\) can matter as much as raw parameter count\.

Foundational medical LLM studies established strong performance on standardized QA and licensing style benchmarks, but also documented residual issues in calibration, omissions, and potential harm in free\-text outputs\[singhal2023medpalm\],\[tu2025expertlevel\]\. Recent frameworks such as CLEVER push evaluation beyond label accuracy toward expert\-reviewed quality dimensions \(clinical relevance, factuality, and justification fidelity\), which is especially important for decision\-support claims\[kocaman2025clever\]\.

DiagnosisArena explicitly targets difficult diagnostic reasoning with long\-form case narratives and broad specialty coverage\[zhu2025diagnosisarena\]\. Compared with multiple\-choice medical QA, this benchmark exposes workflow\-sensitive failure modes: missing a key negative, mishandling timeline constraints, or overfitting to a salient but nonspecific clue\.

Retrieval\-augmented generation provides a principled mechanism to inject updatable external knowledge at inference time\[lewis2020rag\],\[mialon2023augmented\], and agentic reasoning paradigms such as ReAct motivate separating planning, evidence access, and decision actions\[yao2022react\],\[schick2023toolformer\]\. In healthcare, this decomposition improves auditability because reviewers can inspect intermediate artifacts rather than only a final narrative\.

A key line of work asks whether strong prompting can close the gap with domain\-specialized models\. Medprompt\-style results suggest that generalist models can achieve competitive medical benchmark performance under carefully designed prompting\[nori2023medprompt\]\. However, for high\-stakes diagnosis, the literature still supports adding structured verification gates and human expert review protocols, not relying on prompting alone\[kocaman2025clever\]\.

Together, these studies motivate the approach in this paper: treat diagnosis as a constrained, auditable workflow that combines structured clinical extraction, supportive retrieval, candidate\-level comparison, and explicit final\-decision guardrails\.

DiagnosisArena is designed to stress diagnostic reasoning by requiring models to interpret long clinical narratives and produce grounded diagnoses\[zhu2025diagnosisarena\]\. Unlike multiple\-choice medical QA benchmarks, it requires accurate extraction of key positives and negatives, temporal reasoning with timeline consistency, pathology and diagnostic test interpretation when present, and differential diagnosis search under uncertainty\. The failure modes are also workflow\-shaped: missing a single negative, anchoring on a salient but non\-specific clue, or inventing a symptom can derail the final diagnosis\. That makes the benchmark appropriate for evaluating*agentic decomposition*rather than only raw knowledge recall, and motivates the pipeline\-based approach described in the following sections\.

![Refer to caption](https://arxiv.org/html/2607.22555v1/deeplens_1-agent-overview-workflow.drawio.png)Figure 1:Overview of the DeepLens Diagnosis Agent workflow and its core reasoning stages\.
## 3Methods: Five\-Stage Diagnosis Pipeline

This section describes the five sequential stages of the Diagnosis Agent pipeline\. Figure[2](https://arxiv.org/html/2607.22555#S3.F2)illustrates the detailed dataflow across stages\.

![Refer to caption](https://arxiv.org/html/2607.22555v1/fig_2.png)Figure 2:Method pipeline used in this study for structured extraction, retrieval support, candidate generation, and final diagnostic selection\.### 3\.1Base Model: JSL Medical Small

The Diagnosis Agent centers on JSL Medical Small, a 7B\-parameter medical language model fine\-tuned for clinical reasoning tasks\[jsl\_medllm\_docs\]\. On standard medical knowledge benchmarks, this model demonstrates strong performance: 89\.4% on MedQA, 89% on Anatomy, 95% on College Biology, and 88\.2% average across nine standard benchmarks\[jsl\_medllm\_docs\]\. This performance positions it as competitive on narrow fact\-recall tasks commonly used to evaluate medical LLMs\.

However, standard benchmarks do not measure diagnostic reasoning under uncertainty\. When deployed as a single\-shot baseline on DiagnosisArena—a benchmark emphasizing long\-form case interpretation, evidence synthesis, and differential diagnosis—JSL Medical Small achieves only 23\.99% accuracy, a gap of approximately 64 percentage points\. This gap is not unique to this model: a larger 32B\-parameter variant from the same model family, which averages 85\.0% on standard benchmarks, drops to 22\.19% on DiagnosisArena\. The dramatic performance cliff highlights a fundamental distinction: strong performance on medical knowledge recall does not translate to diagnostic workflow execution\.

This observation motivates the core design: rather than scaling model parameters, we embed the 7B model in a constrained pipeline that enforces structured reasoning, explicit evidence grounding, and decision verification\. As demonstrated in Section[6\.2](https://arxiv.org/html/2607.22555#S6.SS2), this approach recovers diagnostic accuracy to 60\.14%, a 36\-point improvement from the vanilla baseline and the highest accuracy among smaller models—achieved through workflow discipline rather than parameter scaling alone\[kocaman2025clever\]\.

### 3\.2Clinical Extraction and Fact Lock\-In

The first step converts raw case text into a structured representation that downstream steps treat as*the only source of patient truth*\. This serves as a primary control against hallucination\.

The extraction step produces a compact, fixed set of fields that the rest of the pipeline treats as the*only*patient truth:main\_case\_text\(core narrative\),timeline,demographics,positives,negatives,diagnostics\(one test per item\), andother\_information\(high\-signal context, explicitly not a differential\)\.

In practice, this step enforces strict machine\-parseable formatting, extraction\-only behavior, diagnostics hygiene \(atomic tests, preserved units, deduplication, specificity ordering\), anti\-leakage constraints onother\_information, and compact source\-of\-truth fields to preserve stability across downstream steps\.

The output is validated and made tolerant to small\-model shape mistakes \(e\.g\., list\-like fields returned as strings are coerced into lists\)\. The intent is to keep the workflow on rails: downstream steps always receive the same*kind*of object \(lists stay lists, required fields stay present\), which reduces silent failure modes\.

As a result, the downstream model becomes a*reasoner over a fact table*rather than a generator operating on unstructured prose\.

### 3\.3Patient\-Level RAG and Pattern Triggers

This stage builds two complementary supports for diagnosis: patient\-level RAG to retrieve relevant medical literature for the overall case\[lewis2020rag\]and pattern triggers to retrieve reusable non\-disease\-specific diagnostic heuristics\.

#### 3\.3\.1Patient\-Level RAG in a Supportive Role

This stage retrieves broad medical literature to support the overall case understanding\. The agent generates one short retrieval query from the extracted fact table, querying across Semantic Scholar and the John Snow Labs curated medical knowledge base \(200M\+ scholarly records\) using a hybrid search strategy\. Query\-generation is intentionally strict: no PHI\-like identifiers \(explicit ages, years, locations, names, or institutions\), preference for discriminators over generic symptoms, and no citations/IDs/URLs or paper\-numbering artifacts\.

To protect privacy, the query is rejected if it includes explicit ages \(e\.g\., 50 years old\) or 4\-digit years\. The query is also kept intentionally short so retrieval stays focused on discriminators rather than copying narrative details\.

If LLM query generation fails, a heuristic query builder composes an embedding query from high\-signal tokens in diagnostics/positives/demographics/other\_information/negatives/timeline\. It applies stop\-word filtering and performs conservative “forced token” injection when critical discriminators appear \(e\.g\.,calcification\-\> forcemyocardial ventricular calcification\)\.

Retrieval uses a hybrid search stack that combines \(i\) fast keyword matching for precise medical terms and \(ii\) lightweight embedding similarity to catch paraphrases and mechanism\-level matches\. The results are then re\-ranked to prioritize articles that are both semantically aligned and clinically specific, rather than merely topically adjacent\.

The literature layer is designed to work at web\-scale by querying across Semantic Scholar plus a large curated medical knowledge base maintained by John Snow Labs\. In practice, this means searching across 200M\+ scholarly records and returning high\-quality candidate articles in milliseconds, which makes retrieval usable as a default supportive step rather than an occasional add\-on\.

From the retrieved pool, the system selects a small fixed set of titles\. Selection enforces a stable count and uniqueness, so the downstream evidence note stays comparable across runs\.

Selected abstracts are compiled into a single RAG context note\. Compilation prompts ban disease labels/abbreviations and ban citations/IDs/links/brackets, specifically to prevent citation leakage and paper\-count bias\. The patient\-level context is kept compact so RAG remains supportive rather than dominant\.

#### 3\.3\.2Runtime Pattern Triggers

In parallel with literature RAG, the agent performs fast local semantic retrieval over similar pattern triggers\.

In practice, the pattern trigger library can be continuously enriched using de\-identified clinical notes and past patient cases \(after review\), capturing recurring discriminator patterns that help steer the differential without turning into disease\-name lookup\.

Pattern entries follow a high\-signal template that includes an arrow \(\-\>\), contains anincludekeyword, and is normalized into a one\-line bullet \(prefixed with•\) to keep guidance simple and reusable for smaller models\.

For retrieval and filtering, the retriever pulls a small pool of patterns, normalizes text for deduplication, and keeps a limited set of unique triggers\. Importantly, this stage is designed to be non\-blocking: if pattern retrieval fails, the pipeline continues with no triggers rather than failing the run\.

Representative pattern triggers include the following: Unilateral or segmental distribution with chronic course suggests mosaic/segmental dermatoses\. Spiny or digitate keratoses suggest digitate hyperkeratosis spectrum versus porokeratotic disorders\. Parakeratotic column findings suggest porokeratosis spectrum and adnexal variants\. When hallmark syndrome features are absent, the differential is widened toward non\-syndromic mimics\.

### 3\.4Constrained Candidate Generation

This step produces a short ranked differential diagnosis \(typically 4 candidates\) with compact,*machine\-checkable*structure\.

Each candidate includes \(i\) a diagnosis label, \(ii\) a bounded likelihood score \(1–10\), and \(iii\) a short reason\. This compactness is not cosmetic: it prevents the candidate step from turning into a second uncontrolled reasoning narrative that contaminates later steps\.

Reliability improves when patient facts dominate and RAG remains supportive; output is constrained to diagnosis labels only \(no findings, tests, or vague descriptors\); candidates must be distinct and exclusive \(no synonyms/near\-duplicates or umbrella \+ subtype pairs\); and pattern triggers are used as internal scaffolding without being quoted as a checklist\.

Candidates are normalized \(to reduce shape variance\), scores are clamped into 1–10, and reasons are kept short and consistent\. This matters because later steps rely on clean candidate strings and bounded scores for stable comparisons and tie\-breaks\.

Additional quality gates that improve performance include deduplication across candidates, balanced coverage \(common plausible \+ rare high\-fit when appropriate\), unifying\-diagnosis checks, mandatory practical discriminators per candidate, and strict format\-compliance repair without introducing new facts\.

### 3\.5Evidence Triangulation with Candidate Literature Support

This stage compares candidates against a quote\-anchored evidence memo, optionally supported by a shared candidate\-literature retrieval pool\.

This stage most directly determines diagnostic correctness by forcing explicit matches and mismatches, rather than relying on fluent plausibility\.

#### 3\.5\.1Candidate\-Level Literature Retrieval

Unlike patient\-level RAG \(Step 2\), which retrieves broad background literature about the overall case, candidate\-level RAG is purpose\-built for discriminating among competing diagnoses\. When enabled, the agent generates one candidate\-literature query intended to discriminate among all candidates\. The query criteria explicitly discourage listing all candidate names and allow at most a small number of diagnosis keywords; the aim is to retrieve discriminators, not to directly retrieve a final label\.

Operationally, this step uses the same hybrid retrieval infrastructure \(keyword \+ lightweight embeddings \+ re\-ranking across Semantic Scholar and the JSL medical knowledge base\), but tuned for*discrimination*rather than broad background\. It selects a small fixed number of titles for stability and compiles an evidence note focused on differentiators \(pitfalls, hallmark discriminators, and what would rule each candidate in or out\) rather than disease\-name lookup\.

This shared\-pool design \(single query, single retrieval\) improves coherence and reduces variance compared to per\-candidate fan\-out\.

#### 3\.5\.2Deterministic Quote\-Anchored Evidence Memo

Rather than asking the model to free\-form compare diagnoses, Step 4 produces a strict TEXT memo used by the final decision step\.

Key criteria are deterministic behavior for identical inputs, strict use of extracted patient facts as the source of truth for FOR/AGAINST, quote\-first bullets that begin with exact verbatim facts in double quotes, explicitUnknownwhen no quote exists, and low\-weight use of literature only for test/pitfall questions rather than patient claims\.

As a hard safety gate, after the memo is generated, the agent filters bullets and drops any bullet whose quoted string does not exactly match a normalized extracted patient fact\. This prevents paraphrased or invented evidence from flipping the final decision\.

In practice, the model’s job becomes to*cite the patient facts verbatim*and isolate what is known vs unknown, which reduces run\-to\-run variance and hallucination compared to narrative reasoning\.

### 3\.6Final Diagnostic Decision and Confidence Scoring

The final step selects the single best diagnosis under strict guardrails and outputs a bounded confidence estimate\.

The final decision returns \(i\) the selected diagnosis label, \(ii\) a bounded confidence score \(0–10\), and \(iii\) a short reasoning string\.

The key prompt guardrails are as follows\.selected\_diagnosismust exactly match one candidate string; no new facts/tests/history may be introduced; reasoning must include one or two verbatim patient quotes from source\-of\-truth fields; “compatible with” pathology phrasing is treated as supportive rather than definitive; lesion\-vs\-etiology conflicts favor the underlying disease; and tie\-breaks use higher score first, then earlier candidate\.

Final decision output is validated for missing fields and the confidence score is clamped, enabling robust downstream consumption\.

### 3\.7Deterministic Inference, Validation, and Failure Handling

Two choices reduce variance and increase success rate for small models: deterministic inference configuration \(temperature 0, top\_p 1\.0, no penalties\) and stage\-specific validation gates \(required fields, bounded scores, unique candidates, quote\-anchored evidence\)\.

The pipeline is explicitly resilient: pattern triggers never fail the run \(they return empty on error\), RAG title selection and consolidation have fallbacks \(first\-K selection and summary concatenation\), and core steps \(clinical extraction, candidate generation, and final decision\) remain required while supportive components degrade gracefully\.

The final answer must not contradict any extracted negative; the chosen diagnosis must have the strongest support\-to\-conflict balance based on the quote\-anchored memo \(scores are tie\-break\-only\); and if no candidate is adequate, the system must returninsufficient evidencewith the top missing discriminator\(s\) rather than fabricate\.

## 4Distinctive Aspects of the Diagnosis Agent

The largest gains did not come from a single prompt\. They emerged from reframing diagnosis as a disciplined workflow\. The underlying ideas are simple, but operationally consequential\.

### 4\.1Separation of Fact Extraction and Medical Reasoning

Instead of asking a model to perform diagnosis in one pass, the method separates extraction from reasoning\.

First, the model rewrites the case into a clean set of clinical facts: positives \(what is present\), negatives \(explicitly absent findings\), diagnostics \(labs, imaging, histopathology\), timeline \(what changed over time\), and demographics/key history\.

Downstream steps are then forced to reason over this fact table rather than over a remembered \(and often distorted\) version of the narrative\. This prevents a common failure mode where the model*approximately*recalls the case and then confidently reasons from the wrong premise\.

### 4\.2Retrieval \(RAG\) as Supportive Evidence

Medical literature retrieval is powerful, but it can bias models in subtle ways through citation\-driven anchoring\. The Diagnosis Agent therefore applies a strict principle: RAG provides context and discriminators \(what to check, what could explain the presentation\)\[lewis2020rag\], while the final decision remains grounded in extracted patient facts\.

This makes the system more robust: if retrieval is imperfect or slightly off\-target, the agent still functions because the decision is constrained by the case facts\.

### 4\.3Pattern Triggers as Reusable Clinical Heuristics

Pattern triggers are reusable heuristics surfaced at the moment the model generates the differential\. They act as reasoning scaffolds for search\-space definition, discriminator selection, and ranking by explanatory coverage with minimal assumptions\.

They are deliberately generic \(not disease\-specific prompts\), and they are injected as private guidance: the model is instructed to use them to think, but not to quote them or treat them as a checklist\.

Over time, these triggers can be enriched from de\-identified clinical notes and prior patient cases \(with appropriate governance\), converting recurring high\-signal patterns into reusable heuristics\.

Examples of high\-signal templates include: unilateral/segmental distribution with chronic course \(consider mosaic/segmental dermatoses and linear variants\), spiny/digitate keratoses \(consider digitate hyperkeratosis spectrum and porokeratotic disorders\), parakeratotic column on histology \(consider porokeratosis spectrum and adnexal variants\), and absence of hallmark syndrome features \(widen toward non\-syndromic mimics\)\.

Additional meta\-triggers that consistently improve performance include unifying\-diagnosis checks \(prefer one coherent explanation when appropriate\), two\-track differentials \(common high\-risk plus rare high\-fit options\), test\-driven discriminators for each candidate when available, and near\-duplicate suppression by retaining the more specific variant\.

### 4\.4Evidence Triangulation Through Quote\-Anchored Comparison

A major source of diagnostic error is invented \(or loosely paraphrased\) evidence\. The agent therefore produces a short memo that compares candidates side\-by\-side and ties every claim back to quoted extracted facts\. This reduces hallucination and makes the final decision step more trustworthy\.

To make the comparison operationally useful, the memo makes support and conflict explicit for each candidate and treats missing evidence as a first\-class outcome rather than an implicit omission\. It also enforces a simple discipline: every claim must be anchored to a quoted patient fact \(or be stated as unknown\), which prevents the comparison from becoming a persuasive but ungrounded story\.

### 4\.5Deterministic Inference and Verifiable Constraints

For diagnostic workflows, reliability is typically more important than novelty\. The system therefore prioritizes stable sampling and strict formats\. This does not increase model intelligence, but it reduces fragility, especially under benchmarking and reproducibility requirements\.

Just as important, the workflow is built around verifiable constraints: bounded scores, exact\-match diagnosis selection, and evidence statements that must point back to extracted facts\. This turns reasoning quality into something partially verifiable, and it reduces the chance that a fluent but ungrounded explanation wins simply because it reads well\.

### 4\.6Transparency Through Retrieved Literature Titles

When retrieval is used, the agent returns the titles of the papers that informed the run\. This transparency feature allows reviewers to quickly scan pulled sources, audit retrieval quality, and detect domain drift \(e\.g\., repeated irrelevant specialties\)\.

Separating broad\-context retrieval from candidate\-discrimination retrieval also helps reviewers interpret system behavior by showing whether the system pulled generally relevant background versus papers that were meant to separate close contenders\. When something goes wrong, this makes it easier to decide whether the issue was retrieval drift or reasoning over otherwise reasonable sources\.

### 4\.7Context Minimization and Role Separation

More context is not always better\. Long, noisy context makes it easier to bury key negatives, mix unrelated snippets, or overfit to a single persuasive paragraph\. The agent uses curated, purpose\-built summaries at each step so the model reasons over high\-signal evidence\.

The workflow also keeps*roles*separated: extraction produces the canonical patient fact table; retrieval provides supportive discriminators; and evidence comparison is forced to cite the extracted facts\. This separation reduces the chance that retrieval dominates the narrative or that a single appealing paragraph overwhelms contradictory negatives\.

## 5Why System Design Can Outperform Model Size

Frontier LLMs are highly capable, but they can underperform in diagnosis benchmarks for operational reasons: they over\-explain instead of deciding, skip negatives and anchor early, violate output constraints \(making scoring and post\-processing harder\), or hallucinate tests and missing details\. The Diagnosis Agent shifts the burden from single\-pass generation to reliable workflow execution\.

### 5\.1Small Reasoning Models as Workflow Executors

A small reasoning model tuned for structured clinical extraction, differential diagnosis writing, evidence\-grounded comparison, and constrained outputs can be a better*workflow executor*\. Retrieval then supplies breadth of knowledge the small model does not need to memorize\[lewis2020rag\]\.

This is especially relevant in DiagnosisArena\-style evaluation\[zhu2025diagnosisarena\], where many failures are not knowledge gaps but workflow failures: missing a negative, collapsing the differential into near\-duplicates, or producing an answer that cannot be reliably parsed and scored\. The agent reduces those failure modes by enforcing structure, gating candidate quality, and forcing evidence comparison to quote patient facts rather than rely on fluent plausibility\. A useful conceptual framing is as follows: a small reasoning model provides structured execution, RAG provides broad external coverage, and the agent workflow provides process\-level supervision across steps\.

### 5\.2Throughput as an Operational Advantage of Smaller Models

Diagnosis workflows are naturally multi\-step, so a single one\-shot completion is rarely sufficient\. A fast smaller model makes it practical to run the behaviors that improve reliability: retrieve context, cross\-check grounding, verify that key negatives were preserved, and apply quality gates without turning the interaction into an expensive, slow process\.

In other words, lower latency enables closed\-loop refinement at the system level: detect errors early \(extraction mistakes, retrieval drift, candidate collapse\), correct them, and continue—without turning every request into an expensive, slow interaction\.

### 5\.3Reproducibility and Experimental Discipline

For scientific reporting and benchmarking, reproducibility matters as much as raw performance\. Practical guidelines include keeping sampling stable for benchmark runs, logging step outputs \(extraction, selected titles, final label\) so failures are explainable, and running ablations \(e\.g\., retrieval off, pattern triggers off\) to quantify true contribution\. Reproducibility is not just a sampling setting; it is also a record\-keeping discipline\. Keeping the intermediate artifacts \(facts, candidates, memo, final label\) and the retrieval titles makes it possible to reproduce failures, compare two workflow variants fairly, and run targeted ablations rather than guessing\.

## 6Evaluation

We evaluate on the DiagnosisArena 915\-case benchmark\[zhu2025diagnosisarena\]\. The primary metric is top\-1 diagnostic accuracy against the benchmark reference label\[zhu2025diagnosisarena\]\. To isolate workflow effects, we compare the proposed agentic system against the same underlying medical\-llm\-small\-7b model without the pipeline, in addition to stronger general\-purpose and reasoning\-oriented baselines\. Table[1](https://arxiv.org/html/2607.22555#S6.T1)reports aggregate rankings\.

### 6\.1Evaluation Methodology

Each model’s diagnostic output is evaluated independently by two LLM judges: Claude Sonnet 4\.5 and Gemini 2\.5 Flash\. Each judge compares the model’s output against the benchmark ground\-truth label and renders a binary correctness judgment \(correct or incorrect\)\. The reported accuracy for each model is the average of the two judges’ scores\.

A diagnosis is judged*correct*if it matches the ground truth exactly, uses an accepted medical synonym or alternative terminology, or appears as the primary diagnosis within a multi\-diagnosis response\. A diagnosis is judged*incorrect*if it identifies a different condition, targets the wrong organ system or disease category, or omits the primary diagnosis\. All models are evaluated with deterministic sampling \(temperature 0\) to support reproducibility\.

This dual\-judge design reduces single\-judge bias and provides a built\-in reliability check\. Across all evaluated models, inter\-judge agreement is 86–87%, and both judges produce consistent model rankings, which supports the validity of the evaluation protocol\.

### 6\.2Results

Table 1:Top\-1 diagnostic accuracy on DiagnosisArena \(average of Claude Sonnet 4\.5 and Gemini 2\.5 Flash judges\)\. The DeepLens Diagnosis Agent achieves 60\.14% accuracy using harnessing with JSL Medical Small \(7B\), the highest among small and medium\-sized models\.RankModelAccuracy1JSL DeepLens Diagnosis Agent60\.14%2Gemini 3 Pro Preview†\\dagger50\.97%3Claude Sonnet 4\.550\.44%4deepseek\-reasoner\-speciale42\.90%5Deepseek\-reasoner40\.11%6Deepseek\-chat36\.01%7JSL Medical Small \(vanilla\)23\.99%8JSL Medical Reasoning 32B \(vanilla\)22\.19%9Llama 3\.3 70B Instruct18\.85%10Nemotron 3 30B \(xhigh\)6\.89%
†\\daggerGemini 3 Pro Preview was evaluated on 464 of 915 cases \(50\.7%\) due to API quota limits; its reported accuracy may differ under full evaluation\.

![Refer to caption](https://arxiv.org/html/2607.22555v1/diagnosisarena_results.png)Figure 3:Diagnostic accuracy comparison on DiagnosisArena\. The JSL DeepLens Diagnosis Agent \(7B v2\) achieves 60\.14% accuracy, the highest among small and medium\-sized models, with a 36\-point gain over the vanilla JSL Medical Small 7B v2 baseline \(23\.99%\)\.Per\-judge accuracy for the Diagnosis Agent is 61\.86% \(Claude Sonnet 4\.5 judge\) and 58\.43% \(Gemini 2\.5 Flash judge\), yielding a 60\.14% average\. The 3\.4\-point inter\-judge spread is consistent with the pattern observed across all evaluated models, where the Claude judge scores 1–3 percentage points higher on average than the Gemini judge\.

The Diagnosis Agent achieves the highest accuracy among small and medium\-sized models, with substantially better performance than Gemini 3 Pro Preview \(\+9\.17pp\), Claude Sonnet 4\.5 \(\+9\.70pp\), and deepseek\-reasoner\-speciale \(\+17\.24pp\)\. Notably, the agent achieves this using a 7B\-parameter medical model with retrieval support, rather than a frontier\-scale generalist\.

The internal ablation baseline is especially informative: the same medical\-llm\-small\-7b model increases from 23\.99% \(vanilla\) to 60\.14% when embedded in the Diagnosis Agent workflow\. This \+36\-point delta supports the central claim that performance gains primarily come from workflow constraints—fact lock\-in, retrieval discipline, candidate gating, and evidence triangulation—rather than parameter scaling alone\.

For deployment\-oriented assessment, label\-based accuracy should be complemented by structured expert review protocols \(e\.g\., CLEVER\) to evaluate clinical reasoning quality, safety, and justification fidelity\[kocaman2025clever\]\.

### 6\.3Generalization to Frontier Models: Can Harnessing Improve Frontier Model Failures?

A critical question for scaling this approach is whether workflow constraints can correct failures from frontier models\. While the agent is centered on JSL Medical Small 7B, we conducted an experiment to test whether the same harnessing framework could systematically address errors made by GPT\-5\.2, one of the strongest models in our benchmark set\.

#### 6\.3\.1Methodology

We selected 10 random cases from DiagnosisArena where GPT\-5\.2 failed \(ground truth accuracy = 0 for that sample\)\. We then ran these same cases through the DeepLens Diagnosis Agent pipeline \(with JSL Medical Small 7B as the base model\)\. The goal was to determine whether structured retrieval, staged reasoning, and explicit evidence triangulation could correct failures from a frontier model, independent of the underlying base model choice\.

#### 6\.3\.2Results

The agent succeeded on 8 out of 10 cases where GPT\-5\.2 failed \(Table[2](https://arxiv.org/html/2607.22555#S6.T2)\)\. The systematic success across diverse diagnostic scenarios reveals how harnessing addresses frontier model failure modes:

##### Case 1 – Differential Diagnosis with Negative Evidence

GPT\-5\.2 predicted “Primary Cutaneous Meningioma,” missing the critical neutrophilic inflammatory signal\. The agent correctly identified “Pyoderma gangrenosum” by prioritizing:

- •Negative microbiology \(“Tissue culture negative for bacteria”\)
- •Neutrophil count recovery coinciding with lesion progression
- •RAG\-retrieved evidence linking neutrophil\-mediated ulceration to pyoderma gangrenosum

This exemplifies how retrieval\-augmented reasoning grounds abstract pattern\-matching in quantitative evidence\.

##### Case 2 – Genetic Evidence Integration

GPT\-5\.2 predicted “Pyoderma gangrenosum,” anchoring on surface features \(ulceration\)\. The agent correctly identified “Familial cerebral cavernous malformation” by:

- •Extracting the KRIT1 pathogenic variant from genetic testing
- •Integrating family history \(father died of intracranial hemorrhage\)
- •Connecting molecular evidence to phenotype \(bluish papules, intracranial hemorrhage, familial recurrence\)
- •Ranking the genetic finding above morphological similarity

This highlights the value of disciplined extraction and evidence triangulation for rare genetic conditions\.

##### Case 3 – Elemental Analysis for Toxicity

GPT\-5\.2 predicted unrelated cutaneous fibrosis\. The agent correctly identified “Gadolinium\-associated plaques” by:

- •Identifying characteristic sclerotic bodies on histopathology
- •Retrieving knowledge that gadolinium accumulates in renal patients
- •Recognizing elemental confirmation \(“Laser ablation inductively coupled plasma mass spectrometry confirmed dermal deposition of gadolinium”\)
- •Distinguishing from systemic fibrosis by lesion localization

##### Case 4 – Organism Identification Overrides Morphology

GPT\-5\.2 predicted “Juvenile Hyaline Fibromatosis\.” The agent correctly identified “Primary cutaneous trichosporonosis” by:

- •Privileging direct microbiological confirmation \(“Culture grew Trichosporon species”\)
- •Using molecular confirmation \(“PCR and DNA sequencing identified Trichosporon mycotoxinivorans”\)
- •Subordinating morphological similarity \(hyaline features\) to direct organism identification

##### Case 5 – Histopathology\-Guided Differential

GPT\-5\.2 confused related dermatological conditions\. The agent correctly identified “Pityriasis rubra pilaris” by:

- •Extracting follicular hyperkeratosis \(present in PRP but atypical for competitors\)
- •Retrieving the characteristic alternating parakeratosis/orthokeratosis pattern
- •Using RAG to distinguish PRP from similar\-appearing psoriasis

Table 2:Agent performance on 10 random GPT\-5\.2 failure cases\. The agent corrected 8 failures, demonstrating that workflow constraints can systematically address frontier model errors\.MetricAgentGPT\-5\.2Correct8/10 \(80%\)0/10 \(0%\)Incorrect2/10 \(20%\)10/10 \(100%\)To verify that the harnessing framework does not degrade performance on cases where frontier models already succeed, we additionally tested 10 random cases where GPT\-5\.2 achieved correct diagnosis\. The agent maintained strong performance on these borderline success cases, demonstrating robustness across the performance spectrum\.

Table 3:Agent performance on 10 random GPT\-5\.2 success cases\. No downgrade: the agent maintains correctness on cases where GPT\-5\.2 already succeeds, indicating the workflow is complementary rather than adversarial\.MetricAgentGPT\-5\.2Correct10/10 \(100%\)10/10 \(100%\)Incorrect0/10 \(0%\)0/10 \(0%\)

#### 6\.3\.3Cost and Latency Implications

The harnessing framework trades immediate latency for reasoning reliability\. At optimized retrieval speeds, the pipeline achieves 24\-second end\-to\-end latency \(Tables[4](https://arxiv.org/html/2607.22555#S7.T4)and[5](https://arxiv.org/html/2607.22555#S7.T5)\), compared to 2\.5 seconds for raw API calls\. This 10x increase is composed of:

- •Patient\-level RAG retrieval: semantic search across medical documents \(optimized: 2–5 seconds\)
- •Candidate\-level RAG retrieval: focused evidence gathering per differential hypothesis \(optimized: 3–8 seconds\)
- •Multi\-stage LLM calls: 5 sequential reasoning stages with intermediate validation \(11 seconds total\)

Latency is context\-dependent\. Batch processing \(e\.g\., retrospective chart review, population screening, EHR enrichment\) amortizes retrieval overhead across cases\. For latency\-sensitive applications, RAG result caching and query parallelization further reduce wall\-clock time\. Current retrieval infrastructure has optimization opportunities that could push latency to 15–18 seconds\.

Cost efficiency, by contrast, favors the harnessed approach: $0\.0072 per case versus $0\.01282 for Gemini 3\.1 Pro or $0\.0110 for Claude Sonnet 4\.5\. Self\-hosted deployment and amortized retrieval infrastructure make this possible\.

#### 6\.3\.4Why Small Tuned LLMs Are a Better Balance

These results illuminate a critical design principle:*model selection should optimize for the task, not just parameter count*\.

- •Medical Domain Tuning: JSL Medical Small 7B achieved 88\.2% accuracy on standard medical benchmarks, a 3\.7x improvement over general\-purpose Llama baselines \(23\.99%\)\. Domain\-specific training gives a small model advantages that frontier generalists cannot replicate without expensive fine\-tuning\.
- •Latency\-Throughput Trade\-off: A 7B parameter model on A100 handles 1\.3M–1\.5M tokens/hour\. This is sufficient for batch workloads \(EHR enrichment, retrospective analysis\) and acceptable for non\-urgent clinical applications\. Frontier models \(Gemini 3\.1 Pro, Claude 4\.5\) offer lower per\-call latency but require cloud APIs and higher per\-token costs at scale\.
- •Auditability: Smaller models are more interpretable\. A 7B model’s internal representations are easier to probe and explain than a 100B\+ frontier model\. This is critical for clinical deployment, where regulators and clinicians expect transparent reasoning\.
- •Cost at Scale: For a 10,000\-case annual diagnostic cohort, the cost difference is material: - –DeepLens Agent: 10,000 cases×\\times$0\.0072 = $72 per year - –Gemini 3\.1 Pro: 10,000 cases×\\times$0\.0128 = $128 per year - –Claude Sonnet 4\.5: 10,000 cases×\\times$0\.0110 = $110 per year Self\-hosted deployment also avoids vendor lock\-in and API reliability risks\.

#### 6\.3\.5When Harnessing Applies: Frontier Models as Baselines

These findings suggest a productive direction for future work: applying the same five\-stage harnessing framework with GPT\-5\.2 as the base model would provide a direct test of whether process constraints can improve frontier model performance beyond their base accuracy\. Such a study would answer whether harnessing is a general technique for strengthening any model, or specific to smaller medical models\. If frontier models can be further improved by harnessing, the cost\-latency trade\-off would shift: higher per\-token costs \(GPT\-5\.2\) but potentially even higher accuracy gains, warranting deployment in safety\-critical scenarios\.

## 7Runtime and Cost Profile

The DeepLens Diagnosis Agent operates with a multi\-step workflow that trades single\-pass latency for reasoning reliability and cost efficiency\. While cloud APIs achieve sub\-3\-second responses, the agent’s structured pipeline enables superior accuracy at competitive cost\.

#### 7\.0\.1Runtime Performance

Table[4](https://arxiv.org/html/2607.22555#S7.T4)and Figure[4](https://arxiv.org/html/2607.22555#S7.F4)show latency across models\. The agent operates at 24\-second end\-to\-end latency on a single case—10x longer than API calls but consistent with batch diagnostic workflows \(EHR enrichment, retrospective review\)\.

Table 4:API response latency per diagnostic case\. DeepLens Diagnosis Agent reflects optimized retrieval speeds; cloud models use public APIs with standard inference\.ModelLatency \(s\)JSL DeepLens Diagnosis Agent \(7B v2\)24\.0DeepSeek\-Reasoner\-Special5\.0Gemini 3\.1 Pro2\.5Claude Sonnet 4\.52\.5![Refer to caption](https://arxiv.org/html/2607.22555v1/runtime_comparison.png)Figure 4:Latency comparison across models\. The DeepLens agent’s 24\-second latency is acceptable for batch and asynchronous diagnostic workflows\. Cloud models offer lower latency \(2\.5–5\.0s\) but at higher per\-token cost\.
#### 7\.0\.2Cost Efficiency

Table[5](https://arxiv.org/html/2607.22555#S7.T5)and Figure[5](https://arxiv.org/html/2607.22555#S7.F5)show inference cost per case\. The agent achieves $0\.0072 per case—lower than Gemini 3\.1 Pro \($0\.0128, 78% higher\) and Claude Sonnet 4\.5 \($0\.0110, 53% higher\)\. Only DeepSeek\-Reasoner\-Special undercuts the agent \($0\.0035\), but at the cost of 17\-point lower accuracy \(42\.90% vs 60\.14%\)\.

Table 5:Inference cost per diagnostic case \(per\-token pricing from public APIs or amortized A100 cost for self\-hosted deployment\)\.ModelCost \($\)JSL DeepLens Diagnosis Agent \(7B v2\)0\.0072DeepSeek\-Reasoner\-Special0\.0035Claude Sonnet 4\.50\.0110Gemini 3\.1 Pro0\.0128![Refer to caption](https://arxiv.org/html/2607.22555v1/runtime_cost_comparison.png)Figure 5:Inference cost per case\. The DeepLens Diagnosis Agent is cost\-competitive with frontier models while delivering superior accuracy\. For annual volumes of 10,000 cases, the agent saves $280–560 compared to cloud APIs\.
### 7\.1Cost\-Accuracy Trade\-off

The DeepLens Diagnosis Agent achieves competitive cost per case \($0\.0072\) while delivering the highest accuracy among small and medium\-sized models \(60\.14%\), demonstrating that workflow discipline can outperform parameter count\. At 24\-second latency, the agent processes cases in near\-real\-time, with cost reflecting 24,000 tokens per case across the full 5\-stage pipeline on A100 infrastructure\. Frontier cloud models achieve lower latency \(2\.5s\) but at 1\.5–1\.8x higher cost \($0\.0110–$0\.0128\) and 9–10 percentage points lower accuracy \(50\.44% for Claude Sonnet 4\.5, 50\.97% for Gemini 3\.1 Pro\)\. The DeepSeek\-Reasoner\-Special model offers lower cost \($0\.0035\) and faster latency \(5\.0s\) but with substantially lower accuracy \(42\.90%, a 17\-point gap from the DeepLens Agent\)\.

For deployment at scale \(e\.g\., 100K cases/year\), the cost\-accuracy trade\-off favors the DeepLens Diagnosis Agent: at $0\.0072 per case, annual costs are $720, compared to $1100–$1280 for cloud APIs—a 35–45% cost reduction while maintaining superior diagnostic accuracy\. This balance of reliability, interpretability, and cost efficiency makes the harnessing approach compelling for healthcare deployment\.

## 8Limitations

Benchmarks compress complex decision\-making into a single gold label, which may not capture clinically reasonable alternative differentials\[zhu2025diagnosisarena\]\. Retrieval can also introduce bias by surfacing misleading or non\-generalizable literature; while the pipeline reduces this risk by treating retrieval as supporting evidence, it cannot remove it entirely\[lewis2020rag\],\[ji2023hallucination\]\.

The system also depends on extraction quality\. Step 1 errors can propagate downstream; schema validation and repair loops reduce this risk but do not eliminate it\. Confidence scores are likewise heuristic unless explicitly calibrated and should be interpreted as internal certainty, not clinical probability\.

Finally, outputs should be treated as decision support, not autonomous diagnosis\. Clinician oversight remains necessary and is consistent with expert\-review best practices in medical LLM evaluation\[kocaman2025clever\]\.

### 8\.1Observed Failure Patterns

Cross\-model coverage analysis reveals a hard ceiling: across all 11 evaluated models, 122 of 915 cases \(13\.33%\) are not correctly diagnosed by*any*model\. These cases are concentrated in rare conditions with complex multi\-factor presentations, including overlap syndromes \(e\.g\., concurrent bullous pemphigoid, anti\-p200 pemphigoid, and anti\-laminin\-γ\\gamma1 pemphigoid\), atypical anatomical sites \(e\.g\., Kaposiform hemangioendothelioma of the internal auditory canal\), and iatrogenic etiologies \(e\.g\., tertiary amine\-induced epithelial keratopathy\)\. This suggests that diagnostic accuracy on rare conditions remains fundamentally limited by current model knowledge, independent of workflow design\.

A separate observation supports the workflow\-over\-scale thesis from a different angle: among the four vanilla DeepLens models evaluated without the agent pipeline, the smallest model \(medical\-llm\-small\-7b, 7B parameters\) achieves 23\.99%, outperforming the 32B \(22\.19%\), 30B \(20\.27%\), and 78B \(14\.15%\) variants\. This inverse scaling pattern suggests that larger parameter counts can dilute domain\-specific knowledge when training data composition shifts toward general\-purpose coverage, and reinforces the rationale for pairing a focused small model with structured workflow constraints rather than scaling parameters alone\.

Different models exhibit complementary diagnostic strengths: each evaluated model \(except the largest DeepLens variant\) solves at least one case that no other model answers correctly\. This complementarity implies that error patterns are partly model\-specific rather than purely difficulty\-driven, which motivates future work on ensemble or routing strategies that combine workflow\-constrained agents with frontier models\.

Finally, outputs should be treated as decision support, not autonomous diagnosis\. Clinician oversight remains necessary and is consistent with expert\-review best practices in medical LLM evaluation\[kocaman2025clever\]\.

## 9Conclusion

We show that agentic workflow design can outperform scaling model size for medical diagnosis tasks\. Structured extraction, disciplined retrieval, pattern\-guided candidate generation, and explicit evidence triangulation yield reliable diagnostic reasoning with a lightweight fine\-tuned model\[jsl\_medllm\_docs\]\.

On DiagnosisArena, the same medical\-llm\-small\-7b model improves from a low\-performing single\-shot baseline to competitive performance once it is embedded in this gated workflow\[zhu2025diagnosisarena\]\. The broader lesson is that for high\-stakes reasoning,*verification gates*\(fact lock\-in, candidate constraints, and quote\-anchored comparison\) often matter as much as raw model capacity\.

In real\-world settings, this workflow is best used as one component inside a multi\-agent medical assistant: it produces a disciplined differential and reviewable evidence memo that other capabilities \(medication safety, guideline retrieval, follow\-up question generation\) can build on, with clinician oversight\.

Future work should evaluate clinical utility beyond top\-1 labels—e\.g\., differential quality, safety, and justification fidelity under structured expert review protocols\[kocaman2025clever\]—and quantify which workflow gates deliver the largest marginal gains\.

## 10Appendix: Example End\-to\-End Run

To make the workflow concrete, we report one representative case with clean, phase\-by\-phase outputs\.

Original input query \(exact\):

Youareanexperienceddiagnosticphysician\.Basedonthefollowingpatientinformation,provideyourmostaccuratediagnosticassessment\.

<case\_information\>

AWhitemaninhis50spresentedwithpainfuloralerosionsthatbegan2yearsbefore\.Medicalhistoryincludedrecurrentrespiratoryinfectionsanddiarrhea\.Apreviousintestinalbiopsyhadshownunspecificfindingsandabsenceofapoptosis\.Hereporteda5\-kgweightlossattributedtoimpairedintakeduetoorallesions\.Hewasbeingevaluatedforamediastinalmassdetectedduringchestradiographyperformedforpneumonia\.

</case\_information\>

<physical\_examination\>

Physicalexaminationrevealedwhitishedematouslacypatcheswitherosionsonthedorsalandlateralaspectsofthetongueandbuccalmucosa\.Skinandnailswerenotinvolved\.

</physical\_examination\>

<diagnostic\_tests\>

\-Biopsy&Immunofluorescence:

\-Lingualbiopsyshowedepidermalacanthosiswithscatterednecrotickeratinocytes,basal\-layervacuolation,andband\-likelymphohistiocyticinfiltrate

\-Directimmunofluorescenceshowedfibrinogendepositionalongbasementmembrane

\-Indirectimmunofluorescencetesting\(usingmonkeyesophagus,salt\-skinsplit,andratbladder\)wasnegative

\-ELISAandimmunoblotanalysisnegativeformultipleantibodies

\-LaboratoryTests:

\-Normalliverfunction

\-NegativeforhepatitisBandC

\-DecreasedIgGserumlevel:465mg/dL\(normal:750\-1600g/L\)

\-NormalIgAandIgMlevels

\-DecreasedtotalBlymphocytecount

\-IncreasedCD8\+TcountwithinvertedCD4:CD8ratio

\-Imaging:

\-Positronemissiontomographyrevealeda10\-cmhypercapturingextrapulmonarymass

\-SurgicalexcisionandhistopathologyofmediastinalmassshowedtypeABthymoma,Masaoka\-Kogastage2A

\-Images:

\-FigureA:Lacy,whiteerosivelesionsonthetongue

\-FigureB:Epidermiswithacanthosis,parakeratosis,necrotickeratinocytes,anddermalbandlymphocyticinfiltrate

\-FigureC:Fibrogendepositionalongthebasementmembrane

\-FigureD:Hypercapturing,heterogeneousmassinrightanteriorhemithorax

</diagnostic\_tests\>

Pleaseanalyzetheinformationandprovideyourdiagnosis\.

Guidelines:

\-Usestandardmedicalterminology

\-Ifmultiplediagnosesapply,listeachonaseparateline

\-Bespecific\(e\.g\.,"Type2DiabetesMellitus"notjust"Diabetes"\)

\-Ifnocleardiagnosiscanbemade,state"Insufficientinformation"

Outputformat:

<thinking\>Useinternalreasoningtodecidediagnosisnamesbutdonotrevealthereasoningprocess\.</thinking\>

<final\_diagnosis\>

Primarydiagnosisname

\[Secondarydiagnosisname,ifapplicable\]

</final\_diagnosis\>

Example:

<final\_diagnosis\>

AcuteAppendicitis

</final\_diagnosis\>

### A\.1 Phase 0: Clinical Fact Extraction

Structured case facts were produced as a compact source\-of\-truth object\. Representative output:

clinical\_extraction

main\_case\_text:AWhitemaninhis50spresentedwithpainfuloralerosions\.\.\.

timeline:Oralerosionsbegan2yearsprior;mediastinalmassdetectedincidentallyduringpneumoniaevaluation\.

demographics:Whitemaninhis50swithrecurrentrespiratoryinfectionsanddiarrhea\.\.\.

positives\(7\):includesoralerosionsandbiopsy/immunofluorescencefindings\.\.\.

negatives\(3\):indirectIFnegative;ELISA/immunoblotnegative;hepatitisB/Cnegative\.\.\.

diagnostics\(7\):includesbiopsy&immunofluorescence,imaging,andlaboratoryentries\.\.\.

other\_information\(3\):includesweightlossandmediastinal\-masscontext\.\.\.

### A\.2 Phase 1: Patient\-Level RAG Support

Patient\-level retrieval used hybrid search over60papers and compiled a compact evidence context used only as supportive input for downstream reasoning\. Relevant titles used in this phase included:*Paraneoplastic pemphigus presenting with a single oral lesion*,*Two Cases of Atypical Bullous Disease Showing Linear IgG and IgA Deposition in the Basement Membrane Zone*, and*A Comparative Study of Direct Immunofluorescence Patterns in Linear IgA Bullous Dermatosis Versus Dermatitis Herpetiformis*\.

### A\.3 Phase 2: Pattern\-Trigger Output

Runtime local retrieval returned10reusable pattern triggers to guide candidate generation \(heuristic scaffolding only\)\. These retrieved triggers are injected into the candidate\-generation prompt as few\-shot guidance\. Example triggers from this run included:

\-Painfulswellingofproximalnailfolds\+lichenoidplaquesinoralmucosa\+positiveanti\-hepatitisCvirusantibody\-\>includebullouslichenplanusofthenails;confirmbiopsyshowinghyperkeratosisanddensebandlikelymphohistiocyticinfiltrate

\-Tensebullaeanderosionswithprogressivetetraparesis\+immunofluorescenceshowinglinearIgGdeposition\-\>includeoverlapofbullouspemphigoid,anti\-p200pemphigoid,andanti\-laminin\-332mucousmembranepemphigoid;confirmhistopathologyandimmunologicalstudies

\-Pruriticblistersandbullaeonhandsandfeet\+elevatedwhitebloodcellcount\+positiveenterovirusPCRfromskinlesions\-\>includeatypicalhand\-foot\-mouthdisease;confirmpunchbiopsyandexcludeotherinfectionswithculturesandserologies

### A\.4 Phase 3: Candidate Generation Output

The model produced4scored candidates:Good Syndrome \(9/10\),Linear IgA Bullous Dermatosis \(7/10\),Mucous Membrane Pemphigoid \(6/10\), andType AB Thymoma \(5/10\)\. Representative output:

batch\_candidate\_generation

candidate\_1:GoodSyndrome\|score=9\|reason:Hypogammaglobulinemia,recurrentinfections,thymoma,andoralerosionswithfibrinogendepositiononimmunofluorescence

candidate\_2:LinearIgABullousDermatosis\|score=7\|reason:Fibrinogendepositionatbasementmembrane,thoughlacksclassicIgAstainingandblisteringmorphology

candidate\_3:MucousMembranePemphigoid\|score=6\|reason:Oralerosionsandlymphohistiocyticinfiltrate,butnoIgG/C3depositionnoted

candidate\_4:TypeABThymoma\|score=5\|reason:Confirmedbyhistopathologyandimaging,butsecondarytounderlyingsyndrome

### A\.5 Phase 4: Differential Analysis Output

A shared micro\-RAG retrieval over50papers \(10selected\) supported a comparative evidence memo over all candidates\.

query:Directimmunofluorescencefibrinogendepositioninterfacedermatitishypogammaglobulinemiathymoma

differentialanalysis:

Thepatientpresentswithacomplexconstellationofsymptomsand

findingssuggestiveofanunderlyingimmune\-mediateddisorderassociated

withathymoma\.Keyfeaturesincludechronicoralerosionswith

characteristichistopathologicalchanges\-\-\-epidermalacanthosis,

necrotickeratinocytes,basal\-layervacuolation,andaband\-like

lymphohistiocyticinfiltrate\-\-\-consistentwithautoimmuneblistering

diseasesuchasparaneoplasticpemphigus\.

Directimmunofluorescencedemonstratesfibrinogendepositionalongthe

basementmembranezone,supportinganautoimmunemechanismtargeting

epithelialstructures\.

\.\.\.\(remainingmemotextcontinuesintheoriginallogoutput\)

Samplerelatedarticletitlesfromthisdifferentialphase:

\-Radiologicchangesinthethymoma\-hypogammaglobulinemiasyndrome\.

\-B\-celllymphopeniaandhypogammaglobulinemiainthymomapatients

\-ThymomaAssociatedwithHypogammaglobulinemia\(Good’sSyndrome\):ReportofaCase

\-Thymomawithchronicdiarrhea

### A\.6 Phase 5: Final Decision Output

Final structured output selectedselected\_diagnosis="Good Syndrome"withconfidence=9/10\. Representative output:

final\_decision

selected\_diagnosis:GoodSyndrome

confidence:9/10

reason:ClinicalfeaturesaligncloselywithGoodSyndrome:thymoma,hypogammaglobulinemia\(IgG465mg/dL\),recurrentinfections,andimmunedysregulation\(invertedCD4:CD8ratio,elevatedCD8\+Tcells\)\.Oralerosionswithfibrinogendepositiononimmunofluorescencesupportanautoimmunemechanismsecondarytothymoma\.Othercandidateslackcriticalcorroborativeevidence:LinearIgABullousDermatosisfailstodemonstrateIgAdeposition,MucousMembranePemphigoidlacksIgG/C3deposition,andTypeABThymomaalonedoesnotexplainsystemicimmunedysfunction\.Verbatimsupport:’Weightlossduetoimpairedintakefromorallesions’\(linkedtoimmunodeficiency\)and’Directimmunofluorescenceshowingfibrinogendepositionalongbasementmembrane’\(supportsautoimmuneetiology\)\.

### A\.7 Run Summary

total\_llm\_calls=10,unique\_candidates=4\. The final diagnosis was selected from explicit candidates using quote\-anchored evidence constraints\.

## References