The Large Cancer Assistant (LCA): A Model-Agnostic Orchestration Framework for Scalable Clinical Decision Support in Oncology

Basarab Matei; Ghassen Marrakchi

REVIEW 3 major objections 6 minor 19 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

Swap any AI model, keep the same clinical routing

2026-07-08 02:42 UTC pith:V35P56CT

load-bearing objection A well-formalized orchestration spec with a sound but near-tautological impermeability proof; the practical question is whether real models can satisfy the interface contract. the 3 major comments →

arxiv 2607.06531 v1 pith:V35P56CT submitted 2026-07-07 cs.AI cs.LG

The Large Cancer Assistant (LCA): A Model-Agnostic Orchestration Framework for Scalable Clinical Decision Support in Oncology

Ghassen Marrakchi , Basarab Matei This is my paper

classification cs.AI cs.LG

keywords clinical decision supportalgorithmic impermeabilitymodel-agnostic orchestrationoncology informaticsmultimodal data standardizationgeometric deep learninginteroperabilityfailure safety

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes the Large Cancer Assistant (LCA), an orchestration framework for oncology clinical decision support that sits between multimodal patient data and black-box AI diagnostic models. The central claim is Algorithmic Impermeability: if two AI models satisfy the same interface contract (input types, output types, and a declared set of failure preconditions), then swapping one for the other leaves the entire orchestration layer—routing decisions, module activation, failure handling, and output schema—completely unchanged. Only the diagnostic content itself differs. The framework is formalized as a 7-tuple architecture with a unidirectional pipeline: data preprocessing, cancer-type routing, diagnostic inference, remedy generation, and output assembly. It introduces an Entry Theory that uses geometric deep learning to standardize heterogeneous clinical inputs (CT scans, pathology slides, clinical notes, blood panels) under a common mathematical representation, while independently tagging each input with a medical signature capturing provenance, usage, and epistemic certainty. The system outputs a Standardized Intermediate Payload (SIP) that deliberately isolates the AI core from hospital IT infrastructure, so changes to electronic medical record systems or interoperability standards never propagate upstream. A proof of concept validated four properties: end-to-end completion with sub-millisecond orchestration overhead, invariant routing projection under model substitution (100% structural equality across swaps), 100% recall on targeted supplementary data requests when inputs are corrupted or missing, and independent multi-protocol execution without cross-contamination.

Core claim

The paper's central result is Proposition 2 (System Impermeability), which states that for any two AI model configurations satisfying the same interface contract, the orchestration-structural projection of the system—activation set, routing decisions, failure handling, and SIP schema—is mathematically identical. The proof works by showing that the preprocessing and routing modules carry no dependence on the AI model parameters, that the failure-precondition sets are part of the interface contract shared by all conforming implementations, and that the SIP schema is determined entirely by protocol-level declarations. This means clinical AI models can be retrained, updated, or replaced without触

What carries the argument

Algorithmic Impermeability (Proposition 2): the guarantee that the orchestration-structural projection is invariant under AI model substitution, provided both models satisfy the same interface contract (Definition 25). The interface contract specifies input types, output types (structural or decisional), and a failure precondition set. The Entry Theory provides the input formalization: a Characterized Entry combines a CS axis (domain, symmetry group, feature space) with a Medical Signature (provenance, usage, epistemic certainty), with these two axes proven mutually independent (Proposition 1). The Cancer Switching Module routes to activated protocols via either deterministic parameterized (

Load-bearing premise

The proof of impermeability assumes that all AI models can be wrapped to satisfy a predefined interface contract, including a declared set of failure preconditions that cleanly separates missing-data failures from internal model crashes. If real-world AI models cannot be reliably wrapped this way, the impermeability guarantee does not hold in practice.

What would settle it

Substitute two real diagnostic AI models (not stubs) that both nominally satisfy the same interface contract but differ in how they handle edge cases—e.g., one model degrades gracefully on partial inputs while the other crashes. If the orchestration-structural projection (routing, failure handling, SIP schema) differs between the two configurations on any input, Proposition 2 is violated. More directly: find a model that cannot be wrapped to cleanly distinguish precondition failures from internal inference failures, and the entire impermeability guarantee collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Clinical AI models in oncology could be upgraded or replaced without revalidating the routing, failure-handling, or output-contract layers of the clinical decision support system, reducing the regulatory and engineering cost of model iteration.
The SIP boundary could allow hospitals to change EMR systems, FHIR versions, or database schemas without modifying any upstream AI orchestration module, since the framework outputs a self-contained payload decoupled from downstream infrastructure.
The failure-safety mechanism (targeted Supplementary Data Requests instead of silent failures) could enforce a human-in-the-loop pathway where missing or corrupted clinical evidence systematically halts the affected protocol rather than producing a speculative diagnosis.
The cancer-agnostic routing design could allow a single orchestration layer to serve multiple cancer types simultaneously, with per-protocol independence preventing cross-contamination between parallel diagnostic pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The impermeability guarantee depends on real-world AI models being reliably wrappable to satisfy the interface contract, particularly the distinction between precondition failures (missing data) and internal inference failures (model crashes). If production models cannot be cleanly wrapped this way, the guarantee is theoretical rather than practical.
The proof of concept uses synthetic stubs rather than real neural networks, so the 100% impermeability result validates the architectural design but not the behavior of the framework under the messier conditions of production clinical AI, where models may partially satisfy contracts or produce ambiguous failure modes.
The V2 probabilistic router, which would auto-detect cancer type from clinical data, is fully specified but not empirically validated—meaning the framework's most clinically ambitious routing capability remains untested.
The Entry Theory's use of geometric deep learning to unify spatial and non-spatial modalities under a common domain-signal representation is mathematically clean but may face practical challenges when clinical data has irregular sampling, missing dimensions, or non-standard acquisition protocols that resist canonical geometric formalization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A well-formalized orchestration spec with a sound but near-tautological impermeability proof; the practical question is whether real models can satisfy the interface contract.

read the letter

Here's the short version: this paper proposes a formal orchestration framework for oncology CDSS that decouples data routing from AI inference. The central impermeability guarantee (Proposition 2) is logically sound but essentially proves what the architecture constructs — the orchestration modules are θ-free by definition, so swapping models that satisfy the same interface contract trivially preserves the routing projection. The real question is whether that contract is satisfiable by actual models, and the PoC can't answer that because it uses synthetic stubs that conform by construction. The stress-test concern about the ⊥/CODE_FAIL distinction is the right one to raise, but I'd push back slightly on its severity: the paper is upfront that it's testing orchestration properties, not clinical accuracy, and the eight canonical cases in Appendix H (including Case 8, the CODE_FAIL path) are well-specified even if only four are empirically exercised. The gap is acknowledged, not hidden. What's genuinely new and well-executed: the Entry Theory applying GDL formalism to standardize heterogeneous clinical inputs (CT volumes, clinical notes, blood panels) under a unified geometric representation is a clean and legitimate extension. The 7-tuple architecture is internally consistent, the SIP/SDR boundary design is thoughtful, and the code is publicly available. The formalization is thorough — Definitions 1–36, properties P1–P5, eight canonical cases mapped to a JSON schema. This is more specification than discovery, but it's a competent specification. Soft spots, in order of importance: (1) The PoC with N=10 synthetic runs and stub models is thin. It verifies the framework doesn't crash, but doesn't test the load-bearing assumption that real models can be wrapped to faithfully distinguish precondition violations (⊥) from internal inference failures (CODE_FAIL). This is the gap between a valid proof and a useful guarantee. (2) V2 (the probabilistic router) is fully specified but unvalidated — the calibration-dependent routing guarantee is deferred entirely. (3) The impermeability guarantee is weaker than it first reads: even with a satisfiable contract, the framework treats model crashes and successes identically at the orchestration level, which is fine for routing but means the guarantee covers less than a casual reader might assume. Minor: the latency argument (sub-ms orchestration overhead) is made against lightweight stubs, not real inference, though the authors acknowledge this. This paper is for systems engineers and informatics researchers thinking about modular clinical AI deployment architectures. It's not for someone looking for a new clinical capability or empirical ML result. It deserves a serious referee who can assess whether the formalization adds enough value over existing modular pipeline patterns to justify publication, and who can push the authors on the contract-satisfiability question. I'd recommend accepting for peer review with a clear expectation that the revision needs at least one real-model integration test.

Referee Report

3 major / 6 minor

Summary. The paper proposes the Large Cancer Assistant (LCA), a model-agnostic orchestration framework for clinical decision support in oncology. The LCA is formalized as a 7-tuple architecture that decouples multimodal data ingestion, clinical routing, and AI inference. The central theoretical contribution is the principle of Algorithmic Impermeability (Proposition 2), which guarantees that the orchestration-structural projection (routing, failure handling, SIP schema) remains invariant under AI model substitution, provided models satisfy a predefined interface contract. The framework includes an Entry Theory based on Geometric Deep Learning to standardize heterogeneous inputs, a Cancer Switching Module (CSM) for routing, and a Standardized Intermediate Payload (SIP) to isolate the system from EMR infrastructure. A Proof of Concept (PoC) with four scenarios validates orchestration properties (nominal flow, impermeability, failure safety, multi-protocol execution) using synthetic data and stub models.

Significance. The paper addresses a genuine systems-engineering gap in clinical AI deployment: the tight coupling of data orchestration and model inference in monolithic architectures. The formalization of Algorithmic Impermeability via Proposition 2 is a clean architectural contribution, and the authors are commendably transparent about the PoC's scope—explicitly stating that it evaluates orchestration behavior, not clinical accuracy, and using synthetic stubs by design. The public code repository and fully specified SIP/SDR JSON schema support reproducibility. The distinction between precondition failure (⊥) and internal inference failure (CODE_FAIL) is a thoughtful design choice for auditability. However, the practical significance is currently limited by the absence of any validation with real models or clinical data, which leaves the realizability of the interface contract untested.

major comments (3)

§3.2.5, Definition 25 + Remark 3: The practical realizability of the ⊥/CODE_FAIL distinction is the load-bearing assumption for Proposition 2's utility. The proof (Appendix B, Step 2) correctly establishes that ⊥-status is identical across conforming implementations because E_⊥^(k) is a declared component of the contract. However, the PoC (S1–S4) uses stubs that trivially conform by construction and does not exercise Case 8 (the CODE_FAIL path) empirically. The paper would benefit from explicitly acknowledging that the contract's decidability assumption—particularly the requirement that wrappers reliably distinguish precondition violations from internal inference failures with real models—is untested, and ideally from sketching how a real model wrapper would implement this distinction (e.g., structural input checks for E_⊥ vs. try/catch for CODE_FAIL).
§4.2, Scenario 2: The impermeability demonstration substitutes Stub A (U-Net-style schema) with Stub B (SegResNet-style schema), both satisfying the same interface contract. Since both stubs are constructed to conform to the contract, the 100% π-equality result is essentially a verification that the codebase correctly implements the projection π, not an empirical discovery. The paper should clarify that S2 is a consistency check of the implementation against Proposition 2, not independent evidence that the proposition holds beyond the contract assumption. This is a framing issue, not a correctness issue.
§3.2.4, Variant V2: The probabilistic auto-detection router is fully specified theoretically (Definitions 18–22) but excluded from the PoC. Since V2 is the more clinically realistic variant (auto-detecting cancer type from data rather than requiring a priori declaration), its absence limits the empirical contribution. The paper should more prominently flag that the framework's practical utility is currently validated only for the deterministic V1 path, and clarify whether V2's calibration-dependent routing guarantee (Definition 21) introduces any additional impermeability risks not covered by Proposition 2.

minor comments (6)

Table 1 lists eight canonical orchestration cases, but the PoC (§4) only covers four scenarios. A brief note mapping S1–S4 to the specific cases tested (e.g., S1→Case 1, S3→Case 4, etc.) would improve traceability.
Figure 2: The shared routing context is denoted as P̂ in the caption but the text uses ̂P (hat over P). Consistent notation would help.
§3.1.1: The claim that the Entry definition 'elegantly unifies spatial and non-spatial modalities' is strong; the trivial symmetry group for tabular data (Table 4) is a degenerate case rather than an elegant unification. Consider softening.
Appendix I: The latency projection equations (Eqs. 3–4) are presented but T_par is not benchmarked. A note that the parallel execution lower bound is unmeasured would prevent misinterpretation.
References: López et al. (2026) is cited as an arXiv preprint with a 2026 date; verify this is not a typo for 2025.
§5.2: The limitations section is thorough but could explicitly mention that the NLG component (lcwm_narrative) is untested for clinical readability or correctness, which is relevant for the SIP's downstream utility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive reading of the manuscript. The referee correctly identifies the paper's core contribution (the formalization of Algorithmic Impermeability and the orchestration framework) and its principal limitation (validation with synthetic stubs only, no real models or clinical data). All three major comments are well-taken: they are framing and scoping clarifications rather than correctness objections. We address each below and confirm the revisions we will make.

read point-by-point responses

Referee: §3.2.5, Definition 25 + Remark 3: The practical realizability of the ⊥/CODE_FAIL distinction is the load-bearing assumption for Proposition 2's utility. The proof (Appendix B, Step 2) correctly establishes that ⊥-status is identical across conforming implementations because E_⊥^(k) is a declared component of the contract. However, the PoC (S1–S4) uses stubs that trivially conform by construction and does not exercise Case 8 (the CODE_FAIL path) empirically. The paper would benefit from explicitly acknowledging that the contract's decidability assumption—particularly the requirement that wrappers reliably distinguish precondition violations from internal inference failures with real models—is untested, and ideally from sketching how a real model wrapper would implement this distinction (e.g., structural input checks for E_⊥ vs. try/catch for CODE_FAIL).

Authors: The referee is correct on both points. First, the decidability of the ⊥/CODE_FAIL distinction with real models is indeed untested by the current PoC. The stubs conform by construction, and while Case 8 is formally specified in Appendix H (Table 8) and the SIP schema (D12) handles it, no PoC scenario exercises the CODE_FAIL path empirically. Second, the referee's suggested wrapper implementation pattern—structural input checks for E_⊥ versus try/catch for CODE_FAIL—is precisely the design we envision. We will add the following to the revised manuscript: (1) an explicit acknowledgment in §5.2 (Limitations) that the contract's decidability assumption is untested with real models, and (2) a concrete remark in §3.2.5 (after Remark 3) sketching the wrapper implementation pattern: E_⊥ checks are structural pre-conditions evaluated before model invocation (e.g., verifying required modality presence, shape conformance, quality thresholds), while CODE_FAIL is emitted when the model raises an exception or returns an invalid output during inference on a structurally valid input. This is a standard exception-handling pattern and does not require the model itself to be aware of the distinction—the wrapper enforces it. revision: yes
Referee: §4.2, Scenario 2: The impermeability demonstration substitutes Stub A (U-Net-style schema) with Stub B (SegResNet-style schema), both satisfying the same interface contract. Since both stubs are constructed to conform to the contract, the 100% π-equality result is essentially a verification that the codebase correctly implements the projection π, not an empirical discovery. The paper should clarify that S2 is a consistency check of the implementation against Proposition 2, not independent evidence that the proposition holds beyond the contract assumption. This is a framing issue, not a correctness issue.

Authors: The referee's characterization is accurate. S2 verifies that the implementation correctly realizes the projection π and that the two conforming stubs produce identical orchestration-structural projections. It is not independent evidence that Proposition 2 holds beyond the contract assumption—the proposition's validity is conditional on the contract being satisfied, and S2 confirms the codebase respects this. We will revise the framing of S2 in §4.2 and the corresponding Discussion paragraph (§5) to state explicitly that S2 is a consistency check of the implementation against Proposition 2 under the contract assumption, not an empirical discovery that the proposition holds independently of that assumption. The current language ('empirically demonstrated algorithmic impermeability') overstates the epistemic status of the result. revision: yes
Referee: §3.2.4, Variant V2: The probabilistic auto-detection router is fully specified theoretically (Definitions 18–22) but excluded from the PoC. Since V2 is the more clinically realistic variant (auto-detecting cancer type from data rather than requiring a priori declaration), its absence limits the empirical contribution. The paper should more prominently flag that the framework's practical utility is currently validated only for the deterministic V1 path, and clarify whether V2's calibration-dependent routing guarantee (Definition 21) introduces any additional impermeability risks not covered by Proposition 2.

Authors: We agree on both the scoping point and the need to address the impermeability question for V2. On the first point: the current PoC is V1-only, and this is stated in §4 and §5.2, but we agree it should be flagged more prominently—specifically in the abstract and the contributions list, so readers do not overestimate the empirical scope. On the second point: V2's calibration-dependent routing (Definition 21) does not introduce additional impermeability risks to Proposition 2, because Proposition 2 concerns the invariance of the orchestration-structural projection under AI model substitution (θ-swap in LCDM/LCRM), and the CSM is θ-free by construction (Step 1 of the proof). V2's routing function h is part of the CSM, not part of the interchangeable AI modules. However, V2 introduces a distinct concern not covered by Proposition 2: if h itself is replaced by a different routing hypothesis h', the activation set P̂ may change, which means the routing projection is NOT invariant under CSM-level substitutions. This is a different kind of substitution than the one Proposition 2 addresses. We will add a remark clarifying this distinction: Proposition 2 covers AI model substitution within the LCDM/LCRM interface contract; V2's routing function is a CSM component and its substitution is outside the scope of the impermeability guarantee. The calibration guarantee (Definition 21) concerns the quality of h's routing decisions, not their invariance under substitution. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; the impermeability proof is straightforward-by-construction but not circular.

full rationale

The paper's central claim (Proposition 2, System Impermeability) is proven by showing that (i) orchestration modules DPM, CSM, LCWM are defined as θ-free, (ii) the ⊥-status of AI modules is determined by the declared precondition set E_⊥ from the interface contract (Definition 25), and (iii) the projection π explicitly discards inference content ỹ (the only θ-dependent component). The proof is near-tautological — 'if you define modules as θ-free and project away θ-dependent content, the projection is θ-invariant' — but it is not circular in the technical sense: the architectural design choices are the inputs and impermeability is the output; they are not the same object renamed. There are zero self-citations in the paper; all references are to external work. No parameters are fitted to data and then 'predicted.' The PoC uses synthetic stubs that trivially conform to the contract, which is a weak empirical test but not circular — it verifies implementation-to-specification conformance, not a fitted quantity. The paper is transparent about the proof being a design property ('by construction'). Score 1 reflects the minor concern that the result is almost definitional, but no circularity pattern from the enumerated list applies.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The framework relies on a priori declared parameters (protocol catalog, activation sets) rather than fitted values. The main axioms are domain assumptions about the applicability of GDL and the ability to wrap AI models in strict contracts. The invented entities (SIP, SDR) are structural components with defined schemas and tested behavior.

free parameters (3)

Protocol Catalog (P) = N/A (declared a priori)
A finite set of clinical protocols declared a priori; not fitted to data but a structural parameter of the system.
V1 Activation Parameter (P_lambda) = N/A (declared a priori)
A direct protocol set declaration for deterministic routing; not computed from data.
V2 Confidence Threshold (tau) = N/A (operational parameter)
A per-protocol threshold for probabilistic activation; declared operationally.

axioms (3)

domain assumption Geometric Deep Learning principles can unify spatial and non-spatial clinical data under a common domain/signal representation.
Section 3.1.1: The Entry Theory formalizes inputs using domains and symmetry groups, assuming this mathematical structure adequately captures clinical data properties.
domain assumption AI models can be reliably wrapped to satisfy strict interface contracts, including the distinction between precondition failures and internal inference failures.
Definition 25 and Remark 3: The impermeability guarantee depends on models conforming to interface contracts, which may not hold for arbitrary real-world models.
ad hoc to paper The orchestration pipeline is a unidirectional DAG with no feedback loops.
Section 3.2.2 (Q3): The framework assumes a strict unidirectional flow, which may not reflect all clinical workflows that require iterative refinement.

invented entities (2)

Standardized Intermediate Payload (SIP) independent evidence
purpose: To serve as the exclusive architectural boundary between the LCA framework and downstream EMR/IT systems.
The SIP is a JSON schema specification (Figure 4, Appendix G) that provides a falsifiable structural contract. Its behavior is tested in the PoC.
Supplementary Data Request (SDR) independent evidence
purpose: To handle failures by halting the pipeline and requesting targeted additional data rather than failing silently.
The SDR mechanism is defined (Definition 34) and its recall is empirically tested in Scenario 3 of the PoC.

pith-pipeline@v1.1.0-glm · 35776 in / 2054 out tokens · 211312 ms · 2026-07-08T02:42:31.562025+00:00 · methodology

0 comments

read the original abstract

- Objective: Multimodal deep learning models in oncology are currently limited by monolithic designs that rigidly couple data ingestion, clinical routing, and artificial intelligence (AI) inference. To address this inflexibility, we propose the Large Cancer Assistant (LCA), a model-agnostic, post-hoc orchestration framework designed for scalable clinical decision support. - Methods: The LCA is mathematically formalized as a 7-tuple architecture grounded in the principle of Algorithmic Impermeability, ensuring the orchestration logic remains strictly independent of underlying black-box AI models. We introduce the Entry Theory, leveraging Geometric Deep Learning (GDL) to standardize multimodal patient data along distinct structural and medical axes. The system dynamically orchestrates data via a Cancer Switching Module and intentionally isolates the core AI execution from volatile hospital IT infrastructures by outputting a Standardized Intermediate Payload (SIP). - Results: A Proof of Concept (PoC) validated the orchestration logic across four technical scenarios. The framework executed a nominal flow with negligible orchestration overhead. It empirically demonstrated algorithmic impermeability by maintaining an invariant routing projection during AI model swaps, and it validated strict failure-safety by achieving a 100\% recall rate in generating targeted Supplementary Data Requests (SDR) under injected data anomalies. Multi-protocol execution capability was also successfully verified. - Conclusion: By structurally decoupling multimodal ingestion from feature inference, the LCA provides a highly adaptable and modular orchestration foundation. The SIP establishes a clear architectural boundary, natively setting the stage for downstream Electronic Medical Record (EMR) interoperability as an independent future paradigm.

Figures

Figures reproduced from arXiv: 2607.06531 by Basarab Matei, Ghassen Marrakchi.

**Figure 1.** Figure 1: Macroscopic ecosystem of the LCA framework, showing the central black box "LCA" with an incoming "Multimodal patient data" stream and three distinct outputs: Diagnosis, Medical Assistance, and a Supplementary Data Request feedback loop. Guo, Long and Deng (2022) or predict post-operative recurrence Wang et al. (2025). Beyond these task-specific tools, several CDSS architectures have been implemented as wo… view at source ↗

**Figure 2.** Figure 2: Internal orchestration graph (unidirectional DAG) showing the pipeline sequence: DPM → CSM → {LCDM, LCRM} → LCWM → SIP. The CSM includes V1 (parameterized) / V2 (auto-detection) annotations. An SDR branch originates from the LCWM. The shared routing context 𝑃̂ is propagated throughout the modules. A foundational design principle of the LCA is Algorithmic Impermeability. The framework imposes strict modula… view at source ↗

**Figure 3.** Figure 3: Structure of the characterized entry and orthogonality of axes, decomposing an entry into the CS axis (Ω, 𝔊, ) and the medical axis 𝜎 = (𝜌, 𝑢, 𝜅). The inset illustrates Proposition 1 (identical CS structure with differing signatures, and identical signatures with differing CS structures). inference content will naturally differ between the two models, but the orchestration layer remains impermeable to th… view at source ↗

**Figure 4.** Figure 4: Condensed SIP/SDR schema (annotated pseudo-JSON representation) showing the block layout: header; input_provenance (entry_refs + 𝜎); csm_execution (variant, 𝑃̂, V1/V2 settings); csm_sdr; protocol_outputs[𝑘] containing individual lcdm_output, lcrm_output, or protocol-level sdr; and finally the lcwm_narrative block. missing or invalid inputs (precondition failures,  (𝑘) ⊥ ). Conversely, if an AI model fail… view at source ↗

**Figure 5.** Figure 5: PoC architecture diagram showing the data pipeline, active LCA modules, synthetic input generation, and reference rule stubs.  (lung) 𝐿𝐶𝐷 , on 𝑁 = 10 identical inputs. The orchestrationstructural projection was invariant in every pair (𝜋-equality 100%) while the diagnostic content differed in every pair (100%; [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: System JSON Output Representation The distinction lcrm_output: null (not invoked) vs lcrm_output.status: "HALTED" (invoked but failed) is semantically important for FHIR: in the first case, no LCRM SDR Task is created. Case 5 LCDM succeeds (NSCLC, 𝑃 = 0.61, 𝜅 = inferred). LCRM emits ⊥: the synthesis policy synth_k of the lung_rads_pipeline protocol only allows therapeutic routing if 𝜅 ≥ suspected; the LCDM… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 4 internal anchors

[1]

and Deng, W

Lin, J. and Deng, W. and Wei, J. and Zheng, J. and Chen, K. and Chai, H. and Zeng, T. and Tang, H. , title =. Journal of Cellular and Molecular Medicine , year =. doi:10.1111/jcmm.70221 , pmid =

work page doi:10.1111/jcmm.70221
[2]

and Wan, Z

Zhang, B. and Wan, Z. and Luo, Y. and Zhao, X. and Samayoa, J. and Zhao, W. and Wu, S. , title =. Frontiers in Pharmacology , year =. doi:10.3389/fphar.2025.1609079 , url =

work page doi:10.3389/fphar.2025.1609079 2025
[3]

and Lin, R

Wang, Z. and Lin, R. and Li, Y. and Zeng, J. and Chen, Y. and Ouyang, W. and Li, H. and Jia, X. and Lai, Z. and Yu, Y. and Yao, H. and Su, W. , title =. Precision Clinical Medicine , year =. doi:10.1093/pcmedi/pbae012 , pmid =

work page doi:10.1093/pcmedi/pbae012
[4]

and Lin, J

Wang, N. and Lin, J. and Li, W. and Lyu, Y. and Jiang, Y. and Ni, Z. and Huang, Q. and Chen, H. and Yan, Q. and Huang, C. , title =. npj Digital Medicine , year =. doi:10.1038/s41746-025-02236-3 , url =

work page doi:10.1038/s41746-025-02236-3
[5]

and Tripathi, A

Waqas, A. and Tripathi, A. and Ramachandran, R. P. and Stewart, P. A. and Rasool, G. , title =. Frontiers in Artificial Intelligence , year =. doi:10.3389/frai.2024.1408843 , url =

work page doi:10.3389/frai.2024.1408843 2024
[6]

and Yang, M

Yang, H. and Yang, M. and Chen, J. and Yao, G. and Zou, Q. and Jia, L. , title =. Briefings in Bioinformatics , year =. doi:10.1093/bib/bbae699 , url =

work page doi:10.1093/bib/bbae699
[7]

and Wang, X

Xiang, J. and Wang, X. and Zhang, X. and Xi, Y. and Eweje, F. and Chen, Y. and Li, Y. and Bergstrom, C. and Gopaulchan, M. and Kim, T. and Yu, K. and Willens, S. and Olguin, F. M. and Nirschl, J. J. and Neal, J. and Diehn, M. and Yang, S. and Li, R. , title =. Nature , year =. doi:10.1038/s41586-024-08378-w , url =

work page doi:10.1038/s41586-024-08378-w
[8]

and Gupta, S

Niecikowski, A. and Gupta, S. and Suarez, G. and Kim, J. and Chen, H. and Guo, F. and Long, W. and Deng, J. , title =. International Journal of Radiation Oncology Biology Physics , year =. doi:10.1016/j.ijrobp.2022.07.894 , url =

work page doi:10.1016/j.ijrobp.2022.07.894 2022
[9]

Studies in Health Technology and Informatics , volume =

Lee, Jaehoon and Kim, JeongAh and Cho, Insook and Kim, Yoon , title =. Studies in Health Technology and Informatics , volume =. 2010 , issn =. doi:, url =

work page 2010
[10]

BMC Medical Research Methodology , volume =

Huser, Vojtech and Rasmussen, Luke V and Oberg, Ryan and Starren, Justin B , title =. BMC Medical Research Methodology , volume =. 2011 , month =. doi:10.1186/1471-2288-11-43 , pmid =

work page doi:10.1186/1471-2288-11-43 2011
[11]

Cardoso, M. J. and Li, W. and Brown, R. and Ma, N. and Kerfoot, E. and Wang, Y. and Murrey, B. and Myronenko, A. and Zhao, C. and Yang, D. and Nath, V. and He, Y. and Xu, Z. and Hatamizadeh, A. and Zhu, W. and Liu, Y. and Zheng, M. and Tang, Y. and Yang, I. and Zephyr, M. and Hashemian, B. and Alle, S. and Darestani, M. Z. and Budd, C. and Modat, M. and V...

work page internal anchor Pith review Pith/arXiv arXiv
[12]

and Lowekamp, B

Beare, R. and Lowekamp, B. and Yaniv, Z. , title =. Journal of Statistical Software , year =. doi:10.18637/jss.v086.i08 , url =

work page doi:10.18637/jss.v086.i08
[13]

ClinicalBERT: Modeling Clinical Notes and Predicting Hospital Readmission

Huang, K. and Altosaar, J. and Ranganath, R. , title =. arXiv preprint arXiv:1904.05342 , year =

work page internal anchor Pith review Pith/arXiv arXiv 1904
[14]

AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation

Ji, Y. and Bai, H. and Yang, J. and Ge, C. and Zhu, Y. and Zhang, R. and Li, Z. and Zhang, L. and Ma, W. and Wan, X. and others , title =. arXiv preprint arXiv:2206.08023 , year =

work page internal anchor Pith review Pith/arXiv arXiv
[15]

Setio, A. A. A. and Traverso, A. and de Bel, T. and Berens, M. S. N. and Bogaard, C. V. D. and Cerello, P. and Chen, H. and Dou, Q. and Fantacci, M. E. and Geurts, B. and Gugten, R. V. and Heng, P. A. and Jansen, B. and de Kaste, M. M. J. and Kotov, V. and Lin, J. Y. and Manders, J. T. M. C. and Sóñora-Mengana, A. and García-Naranjo, J. C. and Papavasilei...

work page doi:10.1016/j.media.2017.06.015 2017
[16]

Computers in Biology and Medicine , volume=

From raw data to research-ready: A FHIR-based transformation pipeline in a real-world oncology setting , author=. Computers in Biology and Medicine , volume=. 2025 , publisher=

work page 2025
[17]

JMIR Medical Informatics , year =

Jung, S and Bae, S and Seong, D and Oh, O and Kim, Y and Yi, B , title =. JMIR Medical Informatics , year =. doi:10.2196/40338 , url =

work page doi:10.2196/40338
[18]

Journal of Medical Internet Research , volume=

Designing clinical decision support systems (CDSS)—A user-centered lens of the design characteristics, challenges, and implications: Systematic review , author=. Journal of Medical Internet Research , volume=. 2025 , publisher=

work page 2025
[19]

An Empirical Analysis of Calibration and Selective Prediction in Multimodal Clinical Condition Classification

An Empirical Analysis of Calibration and Selective Prediction in Multimodal Clinical Condition Classification , author=. arXiv preprint arXiv:2603.02719 , year=

work page internal anchor Pith review Pith/arXiv arXiv

[1] [1]

and Deng, W

Lin, J. and Deng, W. and Wei, J. and Zheng, J. and Chen, K. and Chai, H. and Zeng, T. and Tang, H. , title =. Journal of Cellular and Molecular Medicine , year =. doi:10.1111/jcmm.70221 , pmid =

work page doi:10.1111/jcmm.70221

[2] [2]

and Wan, Z

Zhang, B. and Wan, Z. and Luo, Y. and Zhao, X. and Samayoa, J. and Zhao, W. and Wu, S. , title =. Frontiers in Pharmacology , year =. doi:10.3389/fphar.2025.1609079 , url =

work page doi:10.3389/fphar.2025.1609079 2025

[3] [3]

and Lin, R

Wang, Z. and Lin, R. and Li, Y. and Zeng, J. and Chen, Y. and Ouyang, W. and Li, H. and Jia, X. and Lai, Z. and Yu, Y. and Yao, H. and Su, W. , title =. Precision Clinical Medicine , year =. doi:10.1093/pcmedi/pbae012 , pmid =

work page doi:10.1093/pcmedi/pbae012

[4] [4]

and Lin, J

Wang, N. and Lin, J. and Li, W. and Lyu, Y. and Jiang, Y. and Ni, Z. and Huang, Q. and Chen, H. and Yan, Q. and Huang, C. , title =. npj Digital Medicine , year =. doi:10.1038/s41746-025-02236-3 , url =

work page doi:10.1038/s41746-025-02236-3

[5] [5]

and Tripathi, A

Waqas, A. and Tripathi, A. and Ramachandran, R. P. and Stewart, P. A. and Rasool, G. , title =. Frontiers in Artificial Intelligence , year =. doi:10.3389/frai.2024.1408843 , url =

work page doi:10.3389/frai.2024.1408843 2024

[6] [6]

and Yang, M

Yang, H. and Yang, M. and Chen, J. and Yao, G. and Zou, Q. and Jia, L. , title =. Briefings in Bioinformatics , year =. doi:10.1093/bib/bbae699 , url =

work page doi:10.1093/bib/bbae699

[7] [7]

and Wang, X

Xiang, J. and Wang, X. and Zhang, X. and Xi, Y. and Eweje, F. and Chen, Y. and Li, Y. and Bergstrom, C. and Gopaulchan, M. and Kim, T. and Yu, K. and Willens, S. and Olguin, F. M. and Nirschl, J. J. and Neal, J. and Diehn, M. and Yang, S. and Li, R. , title =. Nature , year =. doi:10.1038/s41586-024-08378-w , url =

work page doi:10.1038/s41586-024-08378-w

[8] [8]

and Gupta, S

Niecikowski, A. and Gupta, S. and Suarez, G. and Kim, J. and Chen, H. and Guo, F. and Long, W. and Deng, J. , title =. International Journal of Radiation Oncology Biology Physics , year =. doi:10.1016/j.ijrobp.2022.07.894 , url =

work page doi:10.1016/j.ijrobp.2022.07.894 2022

[9] [9]

Studies in Health Technology and Informatics , volume =

Lee, Jaehoon and Kim, JeongAh and Cho, Insook and Kim, Yoon , title =. Studies in Health Technology and Informatics , volume =. 2010 , issn =. doi:, url =

work page 2010

[10] [10]

BMC Medical Research Methodology , volume =

Huser, Vojtech and Rasmussen, Luke V and Oberg, Ryan and Starren, Justin B , title =. BMC Medical Research Methodology , volume =. 2011 , month =. doi:10.1186/1471-2288-11-43 , pmid =

work page doi:10.1186/1471-2288-11-43 2011

[11] [11]

Cardoso, M. J. and Li, W. and Brown, R. and Ma, N. and Kerfoot, E. and Wang, Y. and Murrey, B. and Myronenko, A. and Zhao, C. and Yang, D. and Nath, V. and He, Y. and Xu, Z. and Hatamizadeh, A. and Zhu, W. and Liu, Y. and Zheng, M. and Tang, Y. and Yang, I. and Zephyr, M. and Hashemian, B. and Alle, S. and Darestani, M. Z. and Budd, C. and Modat, M. and V...

work page internal anchor Pith review Pith/arXiv arXiv

[12] [12]

and Lowekamp, B

Beare, R. and Lowekamp, B. and Yaniv, Z. , title =. Journal of Statistical Software , year =. doi:10.18637/jss.v086.i08 , url =

work page doi:10.18637/jss.v086.i08

[13] [13]

ClinicalBERT: Modeling Clinical Notes and Predicting Hospital Readmission

Huang, K. and Altosaar, J. and Ranganath, R. , title =. arXiv preprint arXiv:1904.05342 , year =

work page internal anchor Pith review Pith/arXiv arXiv 1904

[14] [14]

AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation

Ji, Y. and Bai, H. and Yang, J. and Ge, C. and Zhu, Y. and Zhang, R. and Li, Z. and Zhang, L. and Ma, W. and Wan, X. and others , title =. arXiv preprint arXiv:2206.08023 , year =

work page internal anchor Pith review Pith/arXiv arXiv

[15] [15]

Setio, A. A. A. and Traverso, A. and de Bel, T. and Berens, M. S. N. and Bogaard, C. V. D. and Cerello, P. and Chen, H. and Dou, Q. and Fantacci, M. E. and Geurts, B. and Gugten, R. V. and Heng, P. A. and Jansen, B. and de Kaste, M. M. J. and Kotov, V. and Lin, J. Y. and Manders, J. T. M. C. and Sóñora-Mengana, A. and García-Naranjo, J. C. and Papavasilei...

work page doi:10.1016/j.media.2017.06.015 2017

[16] [16]

Computers in Biology and Medicine , volume=

From raw data to research-ready: A FHIR-based transformation pipeline in a real-world oncology setting , author=. Computers in Biology and Medicine , volume=. 2025 , publisher=

work page 2025

[17] [17]

JMIR Medical Informatics , year =

Jung, S and Bae, S and Seong, D and Oh, O and Kim, Y and Yi, B , title =. JMIR Medical Informatics , year =. doi:10.2196/40338 , url =

work page doi:10.2196/40338

[18] [18]

Journal of Medical Internet Research , volume=

Designing clinical decision support systems (CDSS)—A user-centered lens of the design characteristics, challenges, and implications: Systematic review , author=. Journal of Medical Internet Research , volume=. 2025 , publisher=

work page 2025

[19] [19]

An Empirical Analysis of Calibration and Selective Prediction in Multimodal Clinical Condition Classification

An Empirical Analysis of Calibration and Selective Prediction in Multimodal Clinical Condition Classification , author=. arXiv preprint arXiv:2603.02719 , year=

work page internal anchor Pith review Pith/arXiv arXiv