Meddollina: Governance-First Clinical AI

• Meddollina is a governance-first system that implements Clinical Contextual Intelligence to ensure safe, clinician-aligned decision support.
• It integrates context structuring, governed reasoning, and language realization to firmly preserve intent and enforce bounded inference.
• Empirical results show consistent performance with explicit uncertainty signaling, deferral under ambiguity, and robust context maintenance.

Meddollina is a governance-first clinical intelligence system designed to instantiate Clinical Contextual Intelligence (CCI) and serve as a continuous intelligence layer within clinical workflows. Unlike generative medical AI systems that treat clinical reasoning as next-token prediction, Meddollina enforces intent preservation, persistent context tracking, bounded inference, and principled deferral under uncertainty. Developed to address the structural limitations of LLM-centric medical AI, Meddollina prioritizes clinician-aligned behavior and responsibility-bound inference, supporting safe, deployable decision support in healthcare environments (S et al., 30 Jan 2026).

1. Clinical Contextual Intelligence (CCI): Foundations and Formalization

Clinical Contextual Intelligence (CCI) is formalized as a distinct behavioral capability class essential for real-world clinical AI deployment. A system exhibits CCI only if it concurrently maintains intent preservation, context persistence, bounded reasoning, responsibility-aware output, and context-bounded truthfulness:

• Intent Preservation: Reasoning remains aligned to the clinician’s explicit goal (e.g., differential diagnosis, management planning) across multi-turn interactions.
• Context Persistence: The system persists an evolving internal state carrying forward explicit facts, identified uncertainties, and open information gaps.
• Bounded Reasoning: Inference is tightly restricted to evidentially justified scopes, with deferral or clarification if contextual support is inadequate.
• Responsibility-Aware Output: Outputs explicitly signal uncertainty, calibrate confidence, and avoid overconfident recommendations in safety-critical cases.
• Context-Bounded Truthfulness: Assertions are supported directly by persistent context or flagged as hypothetical, precluding hallucination.

CCI-oriented decision making is modeled in decision-theoretic terms. Let $C$ denote the clinical context, $A = \{$Answer, Clarify, Defer$\}$ the action set, and $U(C,a)$ the utility of action $a$ under context $C$. A CCI-compliant policy $\pi^\ast$ satisfies: $$\pi^*(C) = \arg\max_{a\in A_\text{allowed}(C)} \mathbb{E}[U(C,a)]$$ subject to credence(answer) $\leq \tau_\text{uncertainty} \Rightarrow a \in \{$Clarify, Defer$\}$, where $A_\text{allowed}(C)$ enforces scope boundaries, and $\tau_\text{uncertainty}$ is the minimum required confidence to proceed without clarification or deferral (S et al., 30 Jan 2026).

2. Meddollina System Architecture and Governance

Meddollina’s architecture is characterized by a strict separation of inference governance from language realization, employing a layered design:

• Context Structuring Layer: Maintains ongoing clinical state $S_C$ with explicit facts (e.g., lab results, clinical history), documented uncertainties, and the clinician’s current intent.
• Governed Reasoning Layer: Applies a governance framework $G$ that constrains inference on $S_C$ to ensure compliance with scope constraints, uncertainty thresholds, and responsibility checks. Inference is delineated and controlled strictly prior to any language output.
• Language Realization Layer: Utilizes a small, parameter-efficient LLM (SLM) to render constrained inferences into clinician-appropriate text, without initiating generative expansion.

The system inverts the generate-then-filter paradigm, enforcing all governance constraints before token-level output. Under ambiguous or underspecified conditions, governance mandates clarifying queries or explicit deferral, never speculative completion. Supported decisions include:

• Deferral: e.g., “I cannot conclude without further data.”
• Clarification request: e.g., “Please specify the patient’s renal function.”
• Maintenance of differential diagnosis rather than premature convergence.

Observable algorithms include uncertainty thresholding, scope enforcement by mapping clinical intent to authorized outputs, and longitudinal consistency checks to detect contradiction or context drift over time. The internal heuristics governing these functions are proprietary, but behavioral outcomes are systematically documented (S et al., 30 Jan 2026).

3. Continuous Intelligence Layer in Clinical Workflows

Meddollina is designed to act as an ongoing, advisory clinical partner rather than an autonomous oracle. All clinical authority is preserved exclusively with human practitioners. Functional features include:

• Context Handoff: Persistent state $S_C$ is synchronized across clinical encounters and with electronic health records or clinician logbooks, maintaining continuity.
• Intent Preservation: The system’s interventions remain constrained to the clinician’s current explicit goal, avoiding unsolicited topic drift.
• Layered Interventions: At any workflow point, Meddollina supplies next-step suggestions, clarification prompts, or defers to a specialist as warranted by role boundaries and case complexity.

By automating context tracking, unresolved question flagging, and systematic enforcement of safety preconditions, Meddollina reduces cognitive burden on clinical staff without diminishing their control over decision-making processes.

4. Experimental Protocols and Evaluation Regimes

Empirical assessment of Meddollina employed a behaviour-first evaluation regime using the MedQuAD benchmark—a suite of 16,412+ heterogeneous medical queries spanning diagnosis, symptom interpretation, treatment planning, genetic conditions, epidemiology, prognosis, and risk assessment. No data filtering or selective curation was applied; the full benchmark acted as a stress test surface.

Four comparator systems were evaluated:

System Class	Core Paradigm
General-purpose generative LLM	Completion-first
Medical-tuned generative model	Knowledge-first
Retrieval-augmented generation	Grounding-first
Meddollina	Governance-first/CCI

The authors emphasize behavioural evaluation over exact-match accuracy, focusing on structure of reasoning (e.g., maintenance of differential), explicit uncertainty management, scope/role appropriateness, behaviourally bounded accuracy (correctness only when justified), and stability under scale (consistency across the full set of queries) (S et al., 30 Jan 2026).

5. Empirical Results and Behavioural Profiles

Quantitative metrics demonstrate Meddollina’s distinct operational profile:

Metric	Generation-Centric AI	Meddollina
Benchmark Completion Rate	$<100\%$ (filtered runs)	$100\%$ (16,412/16,412)
Generation Failure Rate	$>0\%$ under scale	$0\%$
Unsafe Speculative Responses	Common	None observed
Explicit Uncertainty Signaling	Low / inconsistent	Consistent, first-order
Appropriate Deferral Rate	Rare	Frequent when warranted
Premature Diagnostic Commitment	Observed	Not observed
Hallucination under Missing Context	Recurrent	Suppressed
Longitudinal Constraint Drift	Increases with length	Not observed
Scope Boundary Violations	Occasional	None observed
Clinician-Aligned Consistency	Variable	Consistent
Behavioural Degradation at Scale	Present	Absent

Qualitatively, Meddollina requests clarification or defers in underspecified or ambiguous cases, maintains structured differentials where warranted, and prevents context drift or contradiction during extended interactions. All inferences are anchored to $S_C$ or explicitly labeled as hypotheses, suppressing hallucination (S et al., 30 Jan 2026).

6. Discussion and Implications for Medical AI

Meddollina’s results substantiate a paradigm shift away from evaluating medical AI via fluency and accuracy benchmarks alone. Generative scaling, while improving surface-level performance, does not guarantee epistemic restraint, longitudinal coherence, or context-responsivity—requirements central to clinical safety. Domain-tuning and retrieval-augmentation enhance factual accuracy but do not enforce bounded, responsibility-aware behaviour under uncertainty.

By formalizing Clinical Contextual Intelligence and enacting it in a governance-first architecture, Meddollina demonstrates that deployable medical AI must center on explicit reasoning guardrails, persistent context, and principled deferral. The system exemplifies a continuous intelligence model, partnering with clinicians to maintain authority and intent, while rigorously policing the boundaries of algorithmic inference.

A plausible implication is that future regulatory and institutional acceptance of clinical AI will depend less on model parameter count or fluency leaderboards, and more on systematic demonstration of context stewardship, explicit governance, and alignment with clinical workflow safety requirements. The progress metric thus shifts toward behaviourally aligned performance under uncertainty—the operational core of Continuous Clinical Intelligence (S et al., 30 Jan 2026).