An auditable conditional-strategy framework for open-ended decision-making in complex lung cancer

1 Introduction

• Prior clinical AI evaluations often focused on single recommendations or next actions, while agentic systems examined multistep tasks; open-ended choices require showing when each option applies and what changes the decision[1–6].Examples include rare-disease diagnosis with traceable evidence[4], sequential decisions in a sandbox electronic health record[5] and disease-management reasoning across visits[6].Source: §1 · pp. 2–3
• Complex lung cancer denotes a decision point where information and constraints do not support one executable plan, while defensible options, unresolved information or trade-offs remain.It is defined by decision structure at a particular time, not disease stage, rarity, information volume or prognosis; cases with a complete actionable plan and no substantive alternative are outside scope.Source: §1 · p. 3
• Open-ended decisions require plans to connect plausible options to eligibility conditions, decision-changing assessments and actions if a preferred option becomes infeasible or new information emerges.The introduction identifies the gap as not only missing information or recommendations, but also the lack of an explicit, reviewable connection between information, conditions and actions.Source: §1 · p. 3
• The framework responds to the need for teams to turn judgments distributed across sources, specialties and time into a shared plan that makes pathway contingencies reviewable.The introduction notes variation in multidisciplinary-team implementation and deliberation, with clinically meaningful differences in staging, treatment intent and recommendations.Source: §1 · p. 3
• MCE organizes reasonable options, eligibility and exit conditions, unresolved information, safety constraints and contingent next steps in a source-linked document for clinician or multidisciplinary review.It distinguishes known facts from missing or conflicting information, goals and unacceptable risks; tests and treatments remain proposals, with clinical judgment and accountability retained by clinicians.Source: §1 · pp. 3–4
• The evaluation used a purposive corpus enriched for interacting constraints, multidisciplinary case-specific references and physician strategies under unaided, retrieval-reference and MCE-assisted conditions.A separate audit assessed whether MCE connected choices to eligibility, unresolved information and subsequent actions; the corpus attributes were not a validated complexity scale or diagnostic threshold.Source: §1 · p. 4

2.1 Overlapping clinical constraints and case-specific references defined the conditional decision task

• The purposive corpus contained 100 cases, including 74 with unclear guideline recommendations and 63 with interacting factors; the expert reference contained 240 rules for 40 cases.A separate structural audit assessed whether candidates, conditions, unknowns and subsequent actions were connected.Source: §2.1 · pp. 4–5

2.2 MCE linked evidence boundaries and decision-changing unknowns to pathway advancement, de-escalation and fallback

• In an illustrative archived run, verification outcomes governed pathway advancement, de-escalation or continued conservative treatment with bridging, delayed escalation and reassessment.Support for a key premise retained or advanced local treatment; refutation redirected to lower-burden palliation or supportive care, while unresolved uncertainty retained the conservative pathway.Source: §2.2 · p. 6

2.3 MCE support and the expression of case-level clinical requirements in physicians’ final strategies

• Adjusted APAS was 66.27 under M, versus 53.40 under U and 61.05 under R; M–U and M–R differences were 12.87 and 5.22 points, respectively.These results came from 2,250 final strategies produced by 250 physicians across 98 centers.Source: §2.3 · p. 6
• All five content domains showed positive M–U and M–R differences; M3 had the largest descriptive differences at 20.64 and 8.59 points, respectively.Format-aligned MCE outputs also exceeded M-condition physician responses across domains, most strongly for M3 at 29.65 points; they were shorter by 247.5 non-whitespace characters.Source: §2.3 · p. 7

2.4 Conditional clinical-content increment of complete MCE outputs and support from the functional architecture

• Complete MCE outputs had mean APAS of 94.71–97.72 versus 53.16–74.34 for three other models; M3 differences were 28.75–56.25 points in seven of nine combinations.The largest descriptive increment centered on primary and alternative pathways and their eligibility and redirection conditions.Source: §2.4 · pp. 7–8
• Removing the multidisciplinary-prior function reduced APAS by 16.62 points, while removing simulation–critique–repair reduced it by 3.60 points.Complete MCE also exceeded MedGPT base direct generation by 23.82 points and MedGPT-LightRAG by 14.34 points; grouped contrasts do not isolate individual-node effects.Source: §2.4 · p. 8

2.5 APAS has a clinical-interpretation basis, with limits set by relationship closure

• Across 120 outputs, APAS correlated with mean ordinal acceptability at Spearman’s ρ = 0.671; three judges classified 36/40, 30/40 and 40/40 complete MCE outputs as closed across five relationships.At least one judge identified a conflict, conditional misconnection or omission in 12/40 outputs; the audits assess complementary properties and do not replace clinical adjudication.Source: §2.5 · pp. 8–9

• Adjusted APAS differences were 12.87 points versus unaided strategies and 5.22 points versus retrieval-reference strategies.These comparisons were observed in the offline task across 40 fixed cases.Source: §3 · p. 9
• Candidate pathways had the largest descriptive differences, while decision-critical information and safety constraints also showed positive differences.With a shared knowledge base in retrieval-reference and MCE-assisted conditions, the additional content centered on organizing evidence around patient conditions, pathway eligibility, safety constraints and subsequent actions; shorter format-aligned MCE outputs do not establish improved reading or decision efficiency.Source: §3 · pp. 9–10
• MCE is proposed as a shared strategy document for individual preparation and multidisciplinary deliberation, but the study tested only an offline individual-physician task.Evaluation should examine which content clinicians view, adopt, modify, reject or omit, and the burden of review.Source: §3 · p. 10
• Removing multidisciplinary priors produced the largest decrease among targeted ablations, but generation and presentation differences preclude attribution to a single component.Source: §3 · p. 10
• APAS measures expression of encoded requirements, whereas relationship audits assess whether treatment intent, eligibility, action sequence and fallback are correctly connected.Audits found conflicts or omissions in 12 of 40 complete outputs, and physician–AI agreement on individual relationship states was limited; the measures should be used together for human review.Source: §3 · p. 10
• Later-event analyses found decision-relevant content differences but did not show that MCE changed subsequent events or event-group treatment effects.Purposively enriched fixed cases, unrecorded clinician use, evaluator dependence and nonuniform follow-up limit real-world inference. Prospective studies should assess clinician behavior, workflow, safety and clinical consequences before patient-level outcomes.Source: §3 · pp. 10–11

4.2 Case sources and analysis sets

• The 100-case corpus comprised 40 expert-reference cases and 60 source-masked content-evaluation cases, selected for gray-zone attributes rather than a validated complexity scale.Fixed system and evaluation inputs used case summaries and excluded management recommendations, subsequent plans and outcomes; all cases had appeared in development or evaluation, so the corpus was not development-naive.Source: §4.2 · p. 12

4.3 EASS case-specific clinical reference and scoring rules

• A 20-expert panel independently constructed EASS, a 240-rule reference with six rules per case spanning five clinical content domains.Applicable rules were scored 0–2 under fixed rules, and the nine-item CSPR gate treated failures as non-attainment and incomplete majorities as indeterminate; EASS rules were withheld during generation.Source: §4.3 · pp. 12–13

4.4 MCE input and generation of conditional strategies

• MCE used fixed case inputs to represent decision states and goals, retrieve source-linked evidence, construct conditional candidates, simulate scenarios, and review candidates for risks and feasibility.Its five generated candidates were an implementation search budget, not an assertion that only those pathways existed; numeric values supported ranking or checks, not calibrated patient-level treatment effects or risks.Source: §4.4.1 · p. 13, §4.4.2 · p. 14, §4.4.3 · p. 14

4.4.4 Candidate-linked verification and report assembly

• MCE linked verification tasks to candidates and conditional next actions, then checked unresolved items before assembling a Strategy Review Pack.Stopping assessment was workflow control, not formal convergence, global optimality or clinical clearance; proposed governance did not mean review had occurred.Source: §4.4.4 · p. 14

4.5 Audit of strategy relationships in complete MCE outputs

• The relationship audit rated five links in 40 complete MCE packs, while a separate case-variant branch assessed 288 prespecified condition-change requirements across 48 scenarios.Three AI judges rated each relationship dimension; in the case-variant branch, complete closure required every applicable requirement to pass, and the purposive design did not estimate a causal effect or real-world safety.Source: §4.5 · p. 15, §4.14 · pp. 18–19

4.6 Comparators and output processing

• The physician task compared case-summary-only U, retrieval-reference R and MCE-material M across 2,250 responses from 250 physicians, analyzed with a mixed-effects model standardized to the 40 cases.Question sets had nine cases, three per condition; later allocation expansion was not fully reproducible, and the fixed-case estimates were not weighted by clinical frequency or importance.Source: §4.7 · p. 15, §4.16 · p. 19

4.11 Source-masked content evaluation

• Source-masked content modules compared MCE with designated comparators, while separate anchoring and perturbation studies assessed acceptability, defects, rule ratings and responses to reordered or deleted text.The content-evaluation preference unit was a case-comparator module; controlled perturbations used 12 cases and 108 texts, with reordering and content deletion analyzed separately.Source: §4.11 · p. 17, §4.13 · p. 18

4.12 Functional-module ablation across six MedGPT-based configurations

• Six MedGPT-based configurations compared complete MCE with three grouped functional ablations and two reduced-orchestration references in paired fixed-case evaluations.The joint pathway-and-evidence plus simulation ablation did not isolate pathway construction; the design did not estimate individual-node effects or run-to-run stability.Source: §4.12 · p. 17

4.16 Statistical analysis

• CSPR was secondary or exploratory, and its probability difference was not a population-average risk difference across physicians.Fixed-effect predictions were averaged over the case and question-order reference grid with the physician random effect set to zero rather than integrated over its distribution.Source: §4.16 · p. 19
• Primary U/R/M condition effects used mixed-model estimands accounting for repeated physician responses, case heterogeneity and task position; paired case-level differences were supportive.Signed differences were estimated with 95% CIs from 30,000 case-cluster bootstrap samples; response length was not adjusted for as a baseline covariate.Source: §4.16 · p. 20
• Analyses specified separate Holm corrections for APAS and CSPR, distinguished prespecified from post hoc analyses, and reported model diagnostics and agreement measures.Unassigned case-condition cells were not imputed; subsequent-event and other specified strata analyses were unadjusted post hoc, while agreement analyses used measures including kappa, Gwet AC1/AC2 and ICC(A,1).Source: §4.16 · pp. 19–20

4.17 Reproducibility

• Reproducibility linked inputs, generated outputs, ratings, analysis data and figures, while distinguishing recalculation from a new generative-model call.Input versions, software environments and random seeds were recorded; Figure 2 illustrates one traceable run, not a cohort-wide completion rate or patient outcome.Source: §4.17 · p. 20

4.18 Study governance and generative-model reporting

• Study governance de-identified internal cases, restricted access and managed external-model data handling under the approved protocol and institutional requirements.Generative-model roles and reporting details were documented, following TRIPOD-LLM[34] and drawing on DECIDE-AI[30] for early clinical evaluation.Source: §4.18 · pp. 20–21

6 Ethics statement

• The protocol and use of clinical cases were approved by the Peking Union Medical College Hospital Ethics Review Committee (approval no. I-25PJ3487).Source: §6 · p. 21

10 Author contributions

• D.W., Z.H.1, H.S., J.X., and X.X. analyzed the data and drafted the manuscript after a broader group conceptualized the study.Study conceptualization also involved L.Z., S.W.1, and N.L.Source: §10 · p. 22
• D.W., Z.H.1, and colleagues developed case-collection criteria, collected cases, supplied clinical recommendations for system development, and conducted clinician re-evaluations.The contributing group also included Z.Z., Z.B., Y.L.1, Y.L.2, C.G., B.X., K.Z., and N.L.Source: §10 · p. 22
• J.X., S.X., and colleagues developed the scoring strategy, performed blinded assessments, and reviewed clinician re-evaluations.H.S., Z.L., and colleagues developed the system and provided technical support.Source: §10 · p. 22
• D.W., Z.H.1, and colleagues critically reviewed and revised the manuscript, and all authors read and approved its final version.Source: §10 · p. 22

11 Competing interests

• Nine authors are Medlinker employees, and the company developed the MCE system and MedGPT model evaluated in the study.The authors’ roles in study conceptualization, data analysis, manuscript preparation, system development and technical support are described in the Author Contributions statement.Source: §11 · p. 22
• Clinical experts not employed by Medlinker independently conducted the comparative evaluation, and the remaining authors declared no competing interests.Medlinker-employed authors did not participate in scoring-strategy development, blinded assessments or review of clinician re-evaluations.Source: §11 · p. 22

Main Figures

• 12.87 points (95% CI, 11.18 to 14.55) was the adjusted APAS difference for MCE-assisted versus unaided strategies, and 5.22 (3.52 to 6.93) versus retrieval-reference strategies.The main analysis included 2,250 physician–case–condition responses from 250 physicians; estimates were standardized to the fixed 40 cases and balanced task positions.Source: Figure 3 · p. 28

B.5 Format-aligned MCE output

• 29.65 points (95% CI, 23.54 to 35.30) was the largest descriptive within-case M3 difference between format-aligned MCE output and physician responses under M.This comparison assessed candidate clinical pathways; it is distinct from the mixed-model U/R/M condition effects.Source: Figure 3 · p. 28, §B.5 · pp. 34–35

D.1 Cross-model strategy-output comparison

• M3 was the largest descriptive estimate in seven of nine model–judge combinations in the separate cross-model comparison.The comparison used complete MCE and other-model outputs on 40 expert-reference cases, with results reported separately by judge rather than pooled.Source: Figure 4 · p. 29, D Supplementary Methods 4. Supportive… · p. 36, §D.1 · p. 37

Main Figures

• Spearman’s rho between whole-strategy acceptability judgments and APAS was 0.671 (95% CI, 0.581 to 0.752) across 120 outputs.The correlation used the ungrouped mean acceptability score of two physicians and resampled 40 case clusters for its interval.Source: Figure 5 · p. 30

A.1 Expert process

• The expert reference comprised 240 case-specific rules across five content domains, with conditional prerequisites, prohibited actions and fallback boundaries informing APAS and CSPR.Experts resolved unresolved items through further review; reasonable disagreement could leave multiple clinically grounded strategies admissible concurrently.Source: §A.1 · p. 31, §A.2 · p. 32, §A.3 · p. 32

B.1 MCE input, processing and output

• MCE’s Strategy Review Pack linked candidate pathways and evidence to qualifying conditions, verification tasks, reassessment and fallback, while physician adoption or modification remained outside the system boundary.A single illustrative run included 42 source-linked CEUs and a traceable sequence of intermediate objects; it was not a patient outcome or evidence of reproducible wording from a new model call.Source: Figure 2 · p. 27, §B.1 · p. 33

B.3 Audit of five structural relations in complete MCE outputs

• The relationship audit evaluated five links across 40 complete MCE outputs, including whether verification results specified subsequent actions for supportive, refuting or unresolved findings.Three AI judges rated outputs separately; overall closure required all applicable dimensions to be complete, with no critical omission or relationship conflict.Source: Figure 1 · p. 26, §B.3 · pp. 33–34

C.3 Post hoc exploratory analysis of subsequent clinical events

• 14 cases had at least one post hoc component-event record; the record-based categories were not independently adjudicated or validated clinical endpoints.Categories included death, progression or suspected new metastasis, and treatment interruption, non-initiation or health-status deterioration; six cases with unavailable follow-up remained separate.Source: §C.3 · p. 36

D.1 Cross-model strategy-output comparison

• Domain-level APAS used applicable M1–M5 rules with original weights; rankings were descriptive, and fixed-output differences could not be attributed to a single MCE component.Most estimates used 40 cases; Gemini 3.1 Pro Preview M2 estimates used 38 because the rule was inapplicable in two cases.Source: §D.1 · p. 37

D.2.1 First-round allocation and presentation

• Six experts rated 120 A/B modules comparing MCE with MDT-Debate or one of three secondary comparators across five strategy dimensions and an overall judgment.The dimensions covered clinical conflict and goals, patient-state fit, safety, decision-changing verification, and subsequent pathways; experts also selected a principal reason and recorded safety-screen flags.Source: §D.2.1 · p. 37

D.2.2 Independent re-review

• An independent six-clinician panel re-reviewed 60 stratified modules, with reassigned A/B positions and masking to source, first-round ratings, safety-screen fields, and statistical results.The set included 30 MDT-Debate comparisons and ten each for IR-RAG, Direct, and Clinician Plan; each reviewer completed ten modules.Source: §D.2.2 · p. 38

D.2.3 Analysis

• Preference analyses measured agreement on fixed case–comparator modules, not how often MCE was preferred for a dimension.For modules rated in both rounds, overall judgments used a 3 × 3 table, exact agreement, Gwet AC1 and supplementary Cohen κ; dimension-specific judgments used exact agreement, Wilson 95% CIs and unweighted κ.Source: §D.2.3 · p. 38
• Safety-screen flags were summarized by positive and negative agreement, and flags triggered human review rather than confirming a safety event.Intervals did not adjust for correlation from experts rating multiple modules, so agreement applies to this fixed re-review set.Source: §D.2.3 · p. 38

D.3 Functional-module ablation across six configurations

• Six configurations compared complete MCE, targeted functional ablations, direct MedGPT generation and MedGPT-LightRAG, which shared MCE’s knowledge base but omitted its remaining orchestration.Ablations removed multidisciplinary priors, pathway/evidence construction with simulation/critique/repair, or simulation/critique/repair alone; the joint n3–n8 ablation used configuration-limited retrieval.Source: §D.3 · p. 38
• Across 40 cases, GPT-5.5 supplied primary displayed APAS, CSPR and paired within-case comparisons, while Gemini and Claude rankings described configuration-level APAS ranking stability.Contrasts estimate defined bundle removals in fixed cases and single runs; they do not isolate individual-node effects, interactions or run-to-run stability, and output length and organization could change.Source: §D.3 · p. 39

E.1 Overall clinical anchoring

• Two blinded thoracic surgeons independently rated overall clinical acceptability and major clinical defects in 120 outputs sampled from 40 case clusters.Acceptability was the primary anchor and defect presence a supportive diagnostic item; source-specific proportions were descriptive because sampling was balanced and enriched by source, score tier and risk or disagreement.Source: §E.1 · p. 39

E.2 Re-rating EASS rules for complete MCE

• Two physicians independently re-rated 240 EASS rules from 40 complete MCE reports using the original response states and pre-adjudication ratings.Not assessable remained distinct from numeric scores and not applicable; case-level APAS required numeric ratings for all six rules, with rule- and case-level agreement measures specified.Source: §E.2 · pp. 39–40

E.3 Controlled-text perturbation

• Controlled-text perturbations compared 36 original–variant triplets across 12 cases, including content-retaining section-order variants and rule-linked target-content deletions.Three models rated each text three times; nine ratings were averaged, and triplet differences in overall and target-domain APAS used case-cluster bootstrap samples, while partial gates remained supplementary rather than complete CSPR.Source: §E.3 · p. 40

F Supplementary Methods 6. Purposive case variants, relationship closure and evaluator dependence

• The fixed, purposive case-variant analysis locates relationships needing clinical review, not overall MCE error rates, real-world safety-event rates or causal effects of changed conditions.Source: F Supplementary Methods 6. Purposive… · p. 40

F.1 Scenario and CCRR construction

• Sixteen source-case templates generated 48 baseline or modified scenarios, each with six prespecified CCRRs, for 288 requirements spanning pathways, safety constraints and reassessment.Each template had one baseline and two modified scenarios; modifications changed either safety or decision-critical information, or management sequence, branching or reassessment escalation.Source: §F.1 · p. 40
• A fixed three-AI-judge panel assigned five states per requirement; majority determined consensus, with three-way splits marked for clinical review, and AI consensus was not a clinical reference standard.This branch did not use the nine-run U/R/M aggregation; missing calls were separately recorded when the other two judges agreed.Source: §F.1 · p. 40
• Relationship closure was 228 of 287 requirements under conservative denominator handling and 228 of 242 under binary-evaluable handling; neither proportion is an overall performance measure.The conservative denominator retained 45 clinical-review and zero not-assessable requirements without counting them as passing, and excluded one not-applicable requirement. Extended Data Figure 7 shows scenario composition, the template matrix and requirement states.Source: §F.1 · p. 41

F.2 Physician calibration and review cohorts

• Physician calibration used 24 outcome-blind selected scenarios and 144 CCRRs, while targeted review covered flagged or prespecified high-priority requirements, including safety requirements.Two physicians independently assessed the calibration sample; targeted review also used two independent reviewers, with safety and physician-disagreement cases escalated to a third physician who could see earlier opinions.Source: §F.2 · p. 41
• A separate complete re-review covered all 48 scenarios and 288 CCRRs; physician agreement metrics came from that review, while physician–AI comparisons were evaluator-dependence diagnostics.The 24-scenario calibration sample, 161 targeted requirements and 48-scenario re-review retained separate denominators; agreement and judge states are reported in Extended Data Figure 8 and Supplementary Table 8.Source: §F.2 · p. 41

G.1 U/R/M models and standardization

• APAS contrasts were standardized over 40 fixed cases and nine equally weighted task positions using a case- and task-adjusted mixed model with a physician random intercept.Predictions set the random effect to zero, so probability differences are conditional standardized contrasts rather than physician-population-average risk differences.Source: §G.1 · pp. 41–42

G.2 Model estimates for subsequent clinical events

• Event-stratified estimates came from the primary all-case model and were standardized within event target sets; post hoc subgroup analyses were exploratory and did not estimate causal effects.The main text used the 31-case data-quality sensitivity target, with all 34 follow-up cases reported supplementarily; estimates targeted fixed purposive cases and omitted case-sampling and post hoc target-set uncertainty.Source: §G.2 · p. 42

G.3 Model diagnostics and fitted specifications

• Mixed-model formulas, analytic-record counts, convergence, warnings and singularity were recorded; binary models were also checked for separation or weak identification.A case-by-condition binomial model was unstable in the initial GPT-5.5 analysis, with no more specific mechanism identified by available diagnostics.Source: §G.3 · p. 42

G.4 Inter-rater agreement

• Agreement analyses used paired-complete denominators, case- or template-cluster bootstrap intervals, and separate evaluations of ratings, clinical anchoring, perturbations and format-aligned outputs.Repeated ratings were aggregated rather than treated as independent observations; the format-aligned estimate was not a second estimate of the physician-task condition effect.Source: §G.4 · p. 43

G.5 Secondary analyses, multiplicity and sensitivity analyses

• Secondary analyses used Holm adjustment for ten domain comparisons; sensitivity checks varied applicability, input adaptation, response length, judge panels and random-effects structure.Overall APAS and other case-feature analyses were post hoc; length followed condition assignment and was not a baseline covariate in the primary model.Source: §G.5 · pp. 43–44

G.6 Supplementary summaries of relationship closure in case variants

• Relationship-closure summaries distinguished conservative from binary-evaluable requirement denominators and defined complete scenario closure using all six requirements and a paired baseline comparison.Estimates used template-weighted summaries and source-template cluster bootstrap intervals; rule-level and scenario-card agreement were reported separately.Source: §G.6 · p. 44

H Supplementary-material correspondence

• Supplementary materials document the clinical references, design, analytic coverage and object-level summaries, with a de-identified, disclosure-reviewed case-level reproducibility layer planned for the public repository.Original unpublished clinical text, identifying information and free-text follow-up remain subject to privacy, ethics and data-governance requirements.Source: H Supplementary-material correspondence · pp. 45–46

I.1 a. Expert process

• EASS was independently built from fixed case materials, and its content-reference domains were distinct from the main text’s structural-relation audit.EASS experts did not see MCE, comparator or physician outputs, and its rules were not used to generate those outputs.Source: §I.1 · p. 46

I.2 b. Rule fields; J.1 a. APAS operational rules; J.2 b. CSPR gates

• APAS scores expression of applicable case-level requirements, whereas CSPR separately applies nine pathway gates; higher APAS does not imply CSPR passage.APAS rule fields link requirements, applicability, accepted and prohibited actions, fallback actions, weights and scoring; CSPR failure requires failure of an applicable gate.Source: §I.2 · p. 47, §J.1 · p. 47, §J.2 · pp. 47–48

N.2 b. Overall preference and agreement at independent re-review

• 55/60 (91.7%; 95% CI, 81.9%–96.4%) three-category judgments agreed exactly at independent re-review, with Gwet AC1 = 0.910 (0.826–0.982).Unweighted Cohen κ was 0.414 (−0.034 to 0.793); sparse categories affected κ, so exact agreement and Gwet AC1 were emphasized.Source: §N.2 · p. 52

N.3 c. Five clinical-content dimensions in the independent re-review

• D2 patient state and treatment fit showed 44/60 exact agreement (73.3%; 61.0%–82.9%) across rounds for the same 60 fixed text modules.These dimension-specific results measure reproducibility of judgments, not the proportion judged to favor MCE.Source: §N.3 · p. 52

O.1 a. Functional-block and Figure 2 node mapping

• The n6–n8 block assessed candidate benefits, risks, feasibility and dependencies, then repaired or removed candidates; n9 reranked candidates and linked verification objects to subsequent actions.The n10 stopping assessment determined whether the object returned to an earlier stage or was assembled as a Strategy Review Pack.Source: §O.1 · p. 53
• Ablations removed grouped functional blocks because stages within a block operated on dependent intermediate objects, so interventions did not isolate independently acting nodes.Candidate-strategy reranking was retained in both stage-removal configurations; complete retention of every verification behavior was not assumed.Source: §O.1 · p. 53

O.2 b. Configuration matrix

• All six configurations shared the MedGPT base model: Complete MCE was full, three were targeted grouped ablations, and two were reduced-orchestration references.The n3–n8 configuration jointly removed two adjacent blocks and did not isolate the n3–n5 pathway and evidence block; each case–configuration report was generated once.Source: §O.2 · p. 55

O.3 c. Case-level summaries scored by GPT-5.5

• Complete MCE scored mean APAS 91.91 (95% CI, 89.12–94.71), versus 75.29 without multidisciplinary priors, 84.19 without pathway/evidence and review stages, and 88.31 without simulation, critique and repair.Each configuration covered 40 cases; the corresponding case-level success counts were 37/40 (92.5%), 32/40 (80.0%), 37/40 (92.5%), and 37/40 (92.5%), respectively.Source: §O.3 · p. 55

O.4 d. Within-case paired comparisons of complete MCE with other configurations

• Complete MCE exceeded configurations without multidisciplinary priors by 16.62 (95% CI, 11.84–21.40), without simulation, critique and repair by 3.60 (1.72–5.49), MedGPT base direct generation by 23.82 (15.95–31.70), and MedGPT-LightRAG by 14.34 (9.11–19.57) mean APAS points.Differences are complete MCE minus the comparison configuration, based on 40 within-case differences; the supplied passages do not provide the value for the comparison without pathway/evidence and review stages.Source: §O.4 · p. 55

O.5 e. Stability of configuration-level rankings

• Configuration rankings were highly correlated for GPT-5.5 versus Gemini 3.1 Pro Preview (1.000) and GPT-5.5 versus Claude Opus 4.7 (0.943) across six configurations.These correlations describe aggregate ranking stability, not case-level inter-judge agreement or a pooled score.Source: §O.5 · p. 55

P.2 b. Escalation process

• The targeted escalation process used three named AI judges and physician review for specified gaps or disagreements, but did not define a human reference standard or a clinical safety or efficacy endpoint.A three-state split among judges triggered clinical review, and third-physician escalation covered safety requirements and physician disagreements.Source: §P.2 · p. 56

P.3 c. AI-consensus requirement states and descriptive denominators

• Of 287 denominator-eligible requirements, 228 (79.4%) passed under the conservative analysis; among 242 binary-evaluable requirements, 228 (94.2%) passed.The 45 clinical-review requirements counted as not passing in the conservative analysis, while the 14 clear gaps joined passes in the binary-evaluable denominator; one not-applicable requirement was excluded.Source: §P.3 · pp. 56–57

P.4 d. Scenario states by condition category

• For SEQ pathway-sequence scenarios, 3 of 8 reached complete closure, 4 received clinical review, and 1 was a clear gap.Scenario states were jointly derived from six requirements per scenario and use a different unit and decision rule from requirement-level proportions.Source: §P.4 · p. 57

P.5 e. Non-exclusive mechanisms among clear relationship gaps

• Among relationship gaps, pathway-prerequisite unmet was recorded for 8 scenario cards and 6 source templates, while sequence or fallback violation was recorded for 6 and 5, respectively.Mechanism labels were non-exclusive, so their counts do not equal numbers of gap rules, scenarios, or independent clinical events.Source: §P.5 · p. 57

P.6 f. Agreement in complete independent re-review and comparison with AI consensus

• Independent physician re-review showed 67.01% exact agreement and Gwet AC1 0.607 (0.530–0.684) for rules, and 93.75% and 0.892 (0.757–1.000) for scenario cards.Physician–AI rule comparisons showed 39.93% exact agreement and Gwet AC1 0.301 (0.204–0.402) for physician 1 and 39.93% and 0.303 (0.199–0.399) for physician 2; these comparisons diagnose evaluator dependence, not a reference standard.Source: §P.6 · pp. 57–58

P.7 g. Descriptive correspondence between BASE and modified-scenario states

• Across 8 paired SEQ pathway-sequence scenarios, 2 moved toward complete closure, 3 were unchanged, and 3 moved toward a clear gap.These directions describe correspondence within source templates, not causal effects of changed case conditions; model snapshots, prompts, adapters, replicate identifiers, and random conditions were not completely locked across paired scenarios.Source: §P.7 · p. 58

P.8 h. Requirement states by prespecified safety label

• Partial-gate transitions were +0.03 (95% CI, 0.01–0.05) for section-order variants and −0.02 (−0.05 to 0.01) for target-content-deletion variants, not validation of the complete nine-gate CSPR.The transition analysis used the same triplets and applicability denominators; safety-label strata describe requirement states, not clinical safety rates or patient-risk estimates.Source: §P.8 · pp. 58, 60

Q.2 Supplementary Table 9b. Exploratory APAS differences in M1–M5 clinical content domains

• 13.10 APAS points (95% CI, 11.07–15.12) for M1 clinical problem and treatment favored MCE over unaided strategies, and 4.37 (2.32–6.42) favored MCE over retrieval-reference strategies.The Holm-adjusted P values were 6.79 × 10−36 and 7.57 × 10−5, respectively.Source: §Q.2 · p. 61
• 8.59 APAS points (95% CI, 6.35–10.83) for M2 decision-critical information favored MCE over unaided strategies, and 4.88 (2.61–7.15) favored MCE over retrieval-reference strategies.The Holm-adjusted P values were 4.01 × 10−13 and 7.57 × 10−5, respectively.Source: §Q.2 · p. 61
• 20.64 APAS points (95% CI, 18.15–23.13) for M3 candidate clinical pathways favored MCE over unaided strategies, and 8.59 (6.07–11.12) favored MCE over retrieval-reference strategies.The Holm-adjusted P values were 3.60 × 10−58 and 1.61 × 10−10, respectively.Source: §Q.2 · p. 61
• 12.42 APAS points (95% CI, 10.40–14.43) for M4 safety boundaries and risk favored MCE over unaided strategies, and 4.72 (2.68–6.76) favored MCE over retrieval-reference strategies.The Holm-adjusted P values were 1.22 × 10−32 and 2.41 × 10−5, respectively.Source: §Q.2 · p. 61
• M5 reassessment and governance APAS favored MCE by 4.14 points (95% CI, 2.50–5.79) over unaided strategies and 3.47 (1.80–5.13) over retrieval-reference strategies.The Holm-adjusted P values were 3.86 × 10−6 and 7.57 × 10−5, respectively.Source: §Q.2 · p. 61
• Domain APAS included only applicable responses, did not impute non-applicable records, and used Holm adjustment across all ten comparisons.These exploratory analyses describe the clinical-content distribution of the overall APAS difference.Source: §Q.2 · p. 62

b. M1–M5 domain-specific mean APAS

• The supplied M1–M5 domain-specific mean-APAS excerpt lists output, AI judge, and domain headings but provides no estimates.The adjacent judge-specific APAS and CSPR table is identified by title only in the supplied passages.Source: Table 9d · p. 63, b. M1–M5 domain-specific mean APAS · p. 65

c. Within-case M1–M5 differences between complete MCE and other-model strategy outputs

• M3 differences were 40.00 (25.45–54.55) for complete MCE minus GPT-5.4 Gemini 3.1 Pro Preview and 38.75 (24.94–52.56) for complete MCE minus GPT-5.5.Across the reported comparisons, complete MCE-minus-other-model differences were positive across M1–M5; the cited values are the M3 estimates.Source: c. Within-case M1–M5 differences… · p. 67
• Domain APAS retained original rule weights and included applicable rules only, with 40 paired cases except Gemini 3.1 Pro Preview M2 estimates (n=38).Values in panels b and c are means with 95% CIs; intervals were not truncated at the 0–100 scale boundaries.Source: c. Within-case M1–M5 differences… · p. 68
• These supportive within-judge and case comparisons do not isolate a single MCE component or estimate clinical safety.Judges were reported separately without cross-judge pooling, and no between-domain test or multiplicity adjustment was performed.Source: c. Within-case M1–M5 differences… · p. 68

Q.8 Supplementary Table 9h. Data-quality sensitivity and hypothesisgenerating strata for subsequent clinical events

• After three records with prespecified data-integrity concerns were excluded from the standardization target, estimates were standardized over 31 cases while retaining the model fitted to all 40.The model was not refitted after exclusion; six cases with unavailable follow-up remained separate.Source: Table 9f · p. 70, c. Within-case M1–M5 differences… · p. 70
• Post hoc event groups lacked independent clinical adjudication, and the exploratory contrasts were not adjusted for multiplicity or confirmatory effect-modification claims.Absence of a recorded component event did not imply absence of progression or favorable prognosis; zero inclusion or exclusion was not used to confirm domain-specific effect modification.Source: c. Within-case M1–M5 differences… · pp. 70–71
• The subsequent-event comparison included 34 cases with available follow-up, including the three records with data concerns, and was supplementary and exploratory.Domain scores entered the denominator only when the domain was applicable.Source: Table 9g · p. 70, c. Within-case M1–M5 differences… · p. 70

Q.8.1 a. Data-quality and recorded-death strata

• Recorded-death strata contained only four to five cases, and the data-quality analyses were not survival analyses.Because follow-up lacked a uniform index date, observation window, surveillance intensity, and administrative censoring, between-group differences may reflect event-ascertainment opportunity; estimates also omit uncertainty from post hoc target-set selection. In a separate subgroup defined by non-death records of interruption, non-initiation, or deterioration, estimates were directional and no baseline case-level mapping established reasonable foreseeability.Source: §Q.8.1 · p. 73, §Q.8.3 · p. 73

Q.8.2 b. Goal structure and primary clinical-decision scenario

• Primary-plus-fallback cases (n=27) had M–U of 15.05 (95% CI, 12.93–17.16) and M–R of 6.60 (4.46–8.73), versus 7.45 and 1.90 for dual-goal cases (n=12).The dual-goal estimates had 95% CIs of 4.34–10.57 for M–U and −1.26 to 5.07 for M–R; these exploratory strata were not ranked deterministically.Source: §Q.8.2 · p. 73

Q.9.1 a. Overall clinical anchoring

• 74/120 outputs (61.7%) were reported as acceptable, and 71/120 (59.2%) were acceptable to both reviewers.The acceptability analysis also reported reviewer-specific ratings and agreement measures; its source-specific analyses described a fixed enriched sample rather than confirmatory comparisons between sources.Source: §Q.9.1 · p. 74

Q.9.2 b. EASS rule re-rating

• 238/240 rules (99.17%) had adjacent agreement, with weighted Gwet AC2 of 0.842 (95% CI, 0.785–0.893).Case-level ICC(A,1) was 0.694 (95% CI, 0.468–0.835); limits of agreement were descriptive, not clinical-equivalence thresholds.Source: §Q.9.2 · pp. 74–75

Q.10.2 b. Within-domain APAS change for the deleted target content

• For deleted target content, M2 decision-critical information changed by −0.32 (95% CI, −3.41 to 2.13), while M3 candidate clinical pathways changed by −3.70 (−8.02 to 0.00).The accompanying variant-target mapping describes deletion of designated content linked to M2–M5 and physician checks of intended transformations and off-target changes before scoring; this check did not establish a human reference standard.Source: §Q.10.2 · p. 77, §Q.10.3 · p. 77

R.1 a. Operational definitions of gray-zone criteria

• The prespecified gray-zone criteria were non-exclusive and recorded at case level; 78 cases met at least two criteria.Adding documented decision disagreement did not change that count; criteria described purposively assembled case materials, and no baseline-balance test was performed.Source: §R.1 · p. 77, §R.2 · p. 78

R.3 c. Exploratory primary clinical-decision scenarios

• The primary clinical-decision scenarios used a mutually exclusive exploratory classification that did not determine inclusion, grouping, or performance evaluation.Five low-confidence classifications retained their original assignments and remained in the denominator.Source: §R.3 · p. 78

S.1 a. Judge-specific dimensions and overall closure

• Claude Opus 4.7 recorded 40/40 across all seven displayed judge-specific relation and overall-closure entries.The audit covered complete MCE outputs and evaluated relational expression in text; judge observations were reported separately, without voting or averaging.Source: S Supplementary Table 11. Audit of… · pp. 78–79, §S.1 · p. 80

S.2 b. Three-state subsequent actions for verification tasks

• 417/417 observations were complete for subsequent actions and actions after refutation, compared with 416/417 for actions if unresolved.The table reports counts across judge–verification-task observations.Source: §S.2 · p. 81

S.3 c. Outputs with judge-coded defects

• 12 CONFLICT, 2 MISCONNECTION, 2 OMISSION and 0 NOT_EVALUABLE labels were assigned across the 12 anonymized outputs.The output-level labels are shown in the defect table.Source: §S.3 · p. 81

S.4 d. Targeted human checking

• For 10 still-unresolved branches, targeted human checking confirmed the subsequent action in nine and confirmed one omission.Object-level ratings and review records remain governed materials and are not included in the distributable package.Source: §S.4 · p. 81
• The 100-case corpus comprised 40 expert-reference cases and 60 source-masked content-evaluation cases, with no actively excluded cases.The cases were purposively assembled using prespecified gray-zone attributes; internal clinical narratives were restricted and are not publicly released.Source: §S.4 · p. 82

Extended Data Figures

• Extended-data diagnostics covered secondary CSPR and professional-title analyses, anchoring agreement, and score responses to controlled-text perturbations.All six professional-title-group APAS differences were positive with 95% CIs not crossing zero; the interaction test was P = 0.548 and exploratory, while perturbation analysis did not assess clinical validity or patient outcomes.Source: Figure 1 · p. 84, Figure 2 · p. 85, Figure 3 · p. 86
• Source-masked content evaluation reported exact agreement of 40/60, 44/60, 48/60, 56/60 and 48/60 across dimensions D1–D5, respectively.Agreement describes judgments on the same fixed modules, not preference direction or clinical superiority.Source: Figure 4 · p. 87
• At least one judge marked a relationship defect in 12/40 fixed complete MCE outputs; judges were reported separately without voting or performance ranking.The audit covered pathway differentiation, action–premise links, decision-changing unknowns, verification–candidate relations, reassessment/fallback and relationship closure.Source: Figure 5 · p. 88
• Across 48 purposive scenarios, judges classified 13 as completely closed, 23 as requiring clinical review and 12 as containing a clear relationship gap.For prespecified relationship requirements, the conservative descriptive proportion was 228/287 (79.4%) and the binary-evaluable proportion was 228/242 (94.2%); these are not performance or safety rates.Source: Figure 6 · p. 89, Figure 7 · p. 90