Author: Tomasz Jan Gomoła · ORCID 0009-0001-5222-6154 Affiliation: DeepSensi Medical OS — Cognitive Infrastructure Division, DeepSensi PBC, Dover, DE, USA Version: 3.0 — July 2026 Status: Prepared for submission — medRxiv preprint; NEJM AI / npj Digital Medicine Recommended citation: Gomoła, T.J. (2026). "The Hyper Consilium: Integrating Human Clinicians as Scored Nodes in a Fractal AI Diagnostic Swarm." Technical Whitepaper WP-003. DeepSensi Medical OS.
Background. Current human-in-the-loop (HITL) paradigms in clinical AI treat the physician as an external validator — an approval gate positioned after AI processing is complete. This architecture fails to leverage clinical expertise as a cognitive input to the diagnostic reasoning process itself, and provides no mechanism for calibrating the relative reliability of human and AI contributions over time.
Methods. We describe the Hyper Consilium: a fractal cognitive-swarm architecture that integrates human clinicians as mathematically scored, reputation-weighted cognitive nodes operating alongside AI specialist agents within a unified diagnostic pipeline. The architecture implements a unified reputation economy in which AI agents, human physicians, and adversarial-challenge agents share a single reputation registry with type-specific composite influence weighting. The pipeline comprises a set of pre-swarm deterministic barriers, adversarial triadic specialist analysis, multi-round consilium deliberation with dual-pathway (normative and holistic) parallel analysis, a forensic adversarial-challenge protocol, and a continuous red-team swarm. Physician contributions are scored using the same reputation framework as AI agents, with retroactive calibration when post-diagnosis divergence analysis reveals that the physician's judgment outperformed (or underperformed) the swarm's consensus.
Conclusions. The Hyper Consilium is, to our knowledge, the first clinical AI architecture that treats physicians not as external validators but as integral, mathematically calibrated cognitive nodes within a fractal diagnostic swarm — transforming the physician's role from "approval gate" to "cognitive partner" while maintaining the safety guarantees of the DeepSensi Standard (DSS-001) and the multi-layer error bounds established by formal fault tree analysis (WP-001).
Keywords: human–AI teaming; clinical decision support; multi-agent systems; reputation economy; fractal architecture; diagnostic reasoning; physician-in-the-loop; DeepSensi Standard
The dominant paradigm for physician involvement in clinical AI is "human-in-the-loop": the AI generates output, and the physician approves, modifies, or rejects it. While this satisfies the regulatory requirement for human oversight (EU AI Act Article 14), it has fundamental limitations. It is sequential — the physician sees output only after all processing is complete, with no opportunity to inject clinical intuition at earlier stages. It is unweighted — the physician's override carries equal weight regardless of their track record in the specific domain. And it is uncalibrated — there is no mechanism to learn from cases where the physician was right and the AI wrong, or vice versa. The result satisfies the letter of regulation while failing to capture the full value of physician expertise.
The consilium — structured deliberation among multiple specialists to resolve complex cases — has centuries of precedent. The Hyper Consilium translates this tradition into a computational framework with three extensions: (1) the consilium includes both AI specialist agents and human clinicians as deliberation participants; (2) every participant, AI or human, has a mathematically calibrated reputation that weights their contributions; and (3) the deliberation includes structured adversarial challenges designed to expose weaknesses in the emerging consensus.
This paper defines the unified reputation economy governing all node types, describes the multi-round deliberation with dual-pathway analysis and forensic adversarial challenges, and presents the physician-integration architecture that transforms clinical expertise from an external approval function into an integral cognitive input.
Human–AI teaming in medicine has focused on AI-as-assistant and AI-as-second-opinion, both treating physician and AI as separate entities whose outputs are compared after independent processing. Systematic reviews find that combining human and AI judgment can exceed either alone — but only when the integration architecture supports genuine complementarity rather than simple override authority. Reputation systems are well established in distributed computing (PageRank, EigenTrust) for identifying reliable nodes without central authority; their application to clinical AI, treating both AI agents and human physicians as nodes in a reputation-managed cognitive network, has not to our knowledge been previously described. The Hyper Consilium differs from ensemble voting in three respects: it uses adversarial triads within each specialty rather than homogeneous voting; it includes structured challenges that specifically seek to destroy the emerging consensus; and it integrates human judgment as a weighted input alongside AI agents.
Before any cognitive analysis, clinical input passes through deterministic barriers operating entirely outside AI judgment: an input-coherence guard validating clinical biomarker ranges against physiological limits; a document-forensics engine detecting record manipulation; an adversarial-input detector; a cognitive-bias detection barrier that identifies bias in the clinical query itself (anchoring, premature closure, confirmation, diagnosis momentum, availability, framing, false presupposition, input deception) and, when detected, injects both the original frame and an alternative, forcing downstream agents to evaluate both; a case-complexity assessor (a deterministic, sub-millisecond classifier: SIMPLE / MODERATE / COMPLEX / RARE); a knowledge router; and progressive context injection of patient-specific longitudinal data. These barriers are deterministic by design and form the basis of the multi-layer error-probability analysis in WP-001; together with the evidence-verification and post-synthesis tiers, they constitute the 23 independent verification barriers of that analysis, distributed across the orchestration pipeline.
Within each specialty, analysis is conducted by adversarial triads of three cognitively distinct agents: a Lead Analyst (primary assessment), a Creative Challenger / Peer (alternative interpretations from adjacent domains), and a Critical Reviewer / Critic (logical consistency, evidence quality, reasoning soundness). Each specialist is not a prompt template but a multi-layered construct: a domain knowledge profile (scope, criteria, red flags) — the system encompasses 37 specialist profiles across 10 clinical pillars; a persona engine assigning each specialist a distinct communication-style archetype and decision-bias archetype (drawn from defined sets such as rule-out-worst-first, Occam's razor, Bayesian-probabilistic, pattern-recognition, conservative-watchful) plus behavioral rules, simulating the diversity of a real multidisciplinary team; hot-reloadable clinical skill modules (§4.2) injecting physician-contributed knowledge without redeployment; and vendor-agnostic cognitive routing distributing analysis across architecturally independent model families. This last property is not cosmetic redundancy: WP-001 shows that structurally independent model families provide error decorrelation mathematically impossible with copies of the same model; at verification checkpoints, agreement across lineages provides stronger evidence than intra-family consensus. With 37 specialists × 3 triad roles, the system produces up to 111 independent cognitive contributions per session. (Specific persona catalogues, routing policies, and model identities are proprietary.)
Deliberation is an adaptive multi-round convergence procedure. In each round, triads present findings to the full consilium; a formal convergence assessment (set-similarity of diagnostic conclusions across rounds) determines whether additional rounds are needed. The system adaptively runs between three and five rounds based on measured convergence entropy — high-agreement cases converge quickly; high-disagreement cases receive additional rounds.
Two independent analytical pathways execute in parallel. The normative pathway follows evidence-based medicine protocols and guidelines (conservative, guideline-adherent). The holistic pathway employs a systems-biology framework considering cross-system interactions (gut–brain axis, neuroimmune pathways, endocrine–metabolic coupling), environmental factors, and root-cause analysis — but is evidence-grounded, following the same EBM tier hierarchy while searching for explanations guideline-based medicine may not prioritize. A pathway-divergence inspector compares outputs: agreement strengthens confidence; disagreement triggers structured reconciliation, and when divergence is significant, an optional third-path synthesis identifies its source and produces an integrated assessment.
After deliberation and dual-pathway analysis produce a preliminary consensus, the system subjects it to a structured forensic adversarial challenge — a diagnostic interrogation designed to expose weaknesses (What diagnosis was dismissed too quickly? What evidence was under-weighted? What patient-specific factors were missed? What would change with a different age, sex, or ancestry?). A separate cross-vendor verification step re-evaluates the consensus with a model from a different architectural lineage, protecting against systematic biases correlated within a single model family. The entire Hyper Consilium operates through a vendor-agnostic gateway with automatic failover, so conclusions are never dependent on any single provider.
At critical junctures, the meta-cognitive controller may activate a zero-trust cross-verification protocol in which no single component's output is trusted without independent verification from a structurally independent source. Activation is triggered by RARE classification, significant inter-pathway divergence, or meta-cognitive uncertainty signals; when active, the conclusion must survive three independent verification layers (consilium consensus, forensic adversarial challenge, cross-vendor re-evaluation). The conjunction of three structurally independent verifications produces the error-probability bounds analyzed in WP-001.
Continuous adversarial validation is performed by a dedicated three-agent red team, each targeting a specific failure mode: a Miss Hunter (diagnoses the consilium should have considered but did not), an Overconfidence Detector (conclusions whose stated confidence exceeds the evidence), and a Temporal-Blindspot Agent (symptom progressions, medication timing, seasonal variations). Low red-team risk scores force the consilium to reconsider before final output.
The Sentinel system enables real-time physician intervention at any stage — not only final approval. Physicians can inject clinical context at the pre-swarm stage, override individual specialist conclusions with clinical observations, request deeper analysis of specific hypotheses, or halt the pipeline for urgent reasons. A novel capability is autonomous Sentinel summoning: when a patient autonomously initiates a high-tier session, the system identifies and summons relevant Sentinel physicians (by specialty mapping and clinical context) to participate — ensuring expert human oversight is dynamically allocated exactly when and where the patient needs it, even without manual physician initiation. Every physician intervention is recorded in the immutable audit trail with the same granularity as AI actions, enabling precise attribution of each decision element to its source (AI, physician, or deterministic rule).
Physicians can contribute structured clinical skills (diagnostic protocols, treatment guidelines, drug-interaction observations, rare-disease insights, laboratory-interpretation nuances) through a peer-reviewed submission process. Approved skills are registered and become cognitive resources for future sessions; the system tracks citations — instances where a contributed skill influences a conclusion — providing a quantitative measure of each physician's impact on system-wide diagnostic quality.
The central innovation is a unified reputation economy in which AI agents, human physicians, and adversarial-challenge agents share a single registry with type-specific composite influence weighting. Each node's influence weight is a bounded composite of type-appropriate inputs:
The exact coefficient allocations within each composite are proprietary. A malicious-behavior penalty applies escalating reductions to any node (AI or human) exhibiting patterns consistent with adversarial manipulation, ultimately freezing influence at the minimum floor — protecting consilium integrity while preserving the human floor.
When post-diagnosis follow-up data become available (patient outcomes, subsequent findings, referral results), the system computes a divergence score between the original conclusion and the observed outcome. When divergence is significant, retroactive accuracy corrections are applied to every node that participated: nodes that were correct receive a positive adjustment, nodes that were incorrect a (larger) negative adjustment — human Sentinels included, so physicians who consistently provide accurate judgment see their influence increase over time. The economy is thus self-correcting: over many sessions each node's influence converges toward an accurate reflection of its domain-specific accuracy. Meaningful domain-specific calibration requires on the order of a few hundred sessions per node; statistically significant differentiation within a specialty emerges earlier. The system is designed for continuous EHR-integrated deployment, enabling calibration to accumulate through routine use.
When a physician's contributed skill is injected into a session (i.e., influences the analysis), a fixed percentage of session revenue is allocated to a physician royalty pool and distributed by a weighted formula combining originality priority (an inverse function of contribution order, so the first contributor of a novel skill receives the largest share and commodity knowledge converges toward negligible per-physician payout), temporal recency (exponential decay with a defined half-life), and peer validation. (Exact coefficients and half-life are proprietary.) Crucially, royalties are held in escrow behind a Clinical Outcome Calibration Gate: they are released only after retroactive divergence analysis (typically 3–6 months post-diagnosis) confirms the injected skill actually reduced diagnostic latency or correctly altered the outcome; spurious contributions are voided and penalized. This couples economic incentive to verified patient outcomes rather than mere utilization, and structurally eliminates the perverse incentive to inject low-value knowledge to harvest royalties. Because the pool is a fixed percentage of revenue, total payout is mathematically bounded and never exceeds earnings, at any contributor population size. This is, to our knowledge, the first formal economic model for compensating clinical-expertise contributions within an AI diagnostic system — transforming physician knowledge from an uncompensated externality into a measurable, remunerated asset.
Every action — AI analysis, physician intervention, adversarial-challenge outcome, reputation update, royalty calculation — is recorded in a cryptographically secured immutable audit trail. Records are hashed and batched into Merkle trees; each batch's Merkle root is anchored to an external ledger, providing third-party-verifiable proof of existence and integrity. For regulatory review or malpractice litigation, any individual record can be extracted with its Merkle proof, enabling zero-trust point verification without disclosing other records — medical-legal accountability that neither pure-human documentation nor unaudited AI can offer.
Unknown Unknowns Detection. A dedicated detector monitors signal types indicating the pipeline may be operating outside its competence boundary (semantic ambiguity, evidence gaps, specialist disagreement beyond normal variance, temporal inconsistencies, novel symptom combinations), generating one of three recommendations: PROCEED, CAUTION, or INVESTIGATE.
Adaptive depth and latency mitigation. A common critique of large multi-agent systems is latency and cost. The system addresses this through pre-emptive triage — SIMPLE cases bypass the full swarm and execute a low-latency fast-track through a single triad — and a momentum-gated meta-cognitive controller that manages fractal analysis depth using a smoothed (exponential moving average) window of systemic entropy. Depth decisions require sustained entropy rather than reacting to instantaneous spikes, and fractal deepening (re-analysis of high-uncertainty domains) is strictly capped at half the specialists per session, preventing runaway computational cost while ensuring genuine uncertainty receives additional scrutiny.
The structured uncertainty protocol (LIMBO). When the full pipeline — including fractal deepening, additional rounds, and forensic challenge — fails to reach consensus, the system activates LIMBO: a structured non-answer specifying (1) the core of the disagreement, (2) a working hypothesis consistent with the evidence, and (3) the specific investigations that would resolve it. If even the retry fails, the system issues a final response with a reliability score of zero and an explicit "insufficient evidence for definitive verdict." To our knowledge the Hyper Consilium is the only clinical AI with a built-in, formally defined "I don't know" — a capability arguably more important for patient safety than any amount of diagnostic accuracy.
The Hyper Consilium integrates bidirectionally with the AutoResearcher autonomous hypothesis engine (WP-005), the Hypothesis Marketplace (HYPO), and Golden Horizon (WP-006), forming a clinical-intelligence flywheel. Breakthrough hypotheses generated by AutoResearcher are published to HYPO, where they undergo Sentinel peer review within the Hyper Consilium; verified hypotheses enter the medical knowledge base as reusable skills — and because future consilia leverage these verified patterns, the computational cost of diagnosis falls over time. Once a hypothesis is confirmed, it can trigger Golden Horizon to draft compassionate-access and trial proposals, compressing bench-to-bedside translation. Conversely, validated consilium findings feed back into AutoResearcher's belief tracker as high-confidence evidence. The loop: autonomous research generates hypotheses, Sentinel validation tests them, verified knowledge becomes economically rewarded and systematically cheaper to reuse, and breakthroughs bridge into actionable trials.
Hospitals and health systems obtain measurable, auditable diagnostic-quality metrics per session — objective evaluation under the DSS tier framework rather than vendor marketing. Clinicians gain a measurable, portable professional asset: reputation weights that rise with accurate judgment, cited skill contributions, and 37 additional specialist perspectives amplifying (not replacing) their judgment. Patients benefit from breadth (37 specialist perspectives) plus depth (their physician's expertise), subjected to adversarial validation, evidence-quality verification, and a formal uncertainty protocol, with a Merkle-anchored audit trail providing accountability neither pure-human nor pure-AI processes can match. Vendor independence ensures procurement is not locked to any single AI vendor; if a provider changes pricing or discontinues its API, the pipeline continues uninterrupted.
Theoretical diagnostic advantage (illustrative). If a single specialist agent achieves accuracy p for a case type, three cognitively diverse agents analyzing independently give an effective accuracy under an independence assumption of 1 − (1 − p)³ — e.g., 0.90 → 0.999. This calculation assumes full error independence across triad members, a condition approximated but not achieved in practice, since agents share underlying model capabilities; the persona engine and cross-vendor routing are specifically designed to maximize decorrelation, but the figure is an upper bound, not an empirical result. Empirical diagnostic performance is reported separately (WP-001 §5: N = 301 NEJM CPC cases; 86.0% top-1, 93.7% top-3, 0.0% missed-critical after safety-first calibration).
Prospective clinical validation is actively underway; the architecture integrates with EHR systems via FHIR R4 and HL7 v2, supports cloud and air-gapped on-premises deployment, and protects data with strong encryption, automated PII anonymization, and consent-gated processing. Reputation-economy calibration requires substantial longitudinal data (order of a few hundred sessions per physician node). Physician adoption presents a human-factors challenge: clinicians must accept mathematical scoring of their contributions, which may meet resistance in hierarchical clinical cultures; the human floor guarantee and CWPD incentives mitigate but do not eliminate this. Scaling multi-round deliberation carries computational cost and latency, mitigated by the adaptive-depth controller. Cultural variation means reputation weights calibrated in one system may not transfer without recalibration. Independence assumptions in the theoretical advantage (§7) are approximated, not guaranteed. Generalizability across settings, populations, and medical cultures requires evaluation.
The Hyper Consilium is, to our knowledge, the first clinical AI architecture that integrates human clinicians as mathematically scored cognitive nodes within a fractal AI diagnostic swarm. The unified, retroactively calibrated reputation economy transforms the physician's role from external validator to integral cognitive partner. The multi-phase pipeline — deterministic safety barriers, adversarial triadic analysis, multi-round deliberation, dual-pathway architecture, forensic adversarial challenge, and structured uncertainty protocol — provides a cognitive framework exceeding any single AI model or clinician operating alone. Integrated with the DeepSensi Standard (DSS-001), the fault-tree safety analysis (WP-001), and the AutoResearcher–HYPO engine (WP-005), it positions diagnosis, research, and clinical expertise as a unified, reputation-governed, safety-verified system. Certain components (the unified reputation economy, the forensic adversarial-challenge protocol, the cognitive-bias detection barrier, and the LIMBO uncertainty protocol) are the subject of patent applications.
Tomasz Jan Gomoła is the founder and System Architect of DeepSensi PBC, which developed the Hyper Consilium architecture. Certain components are the subject of patent applications. No external funding was received.
© 2026 Tomasz Jan Gomoła / DeepSensi PBC. The Gomola Framework and DeepSensi Standard are open, royalty-free (attribution required). The reference implementation (DeepSensi Medical OS) is proprietary. Patent applications pending.