Chapter 18: Open Questions and the Frontier

The Central Empirical Gap

The 9 \times 9 covariance matrix \Sigma is the framework’s central empirical object. Chapter 17 proposed experimental designs for estimating it — retrospective ethics consultation coding and prospective discrete choice experiments. These designs are feasible, but they have not been executed. Until \Sigma is estimated, the framework’s quantitative predictions remain theoretical.

Specific Open Questions

Question 1: Is \Sigma universal or context-dependent?

The framework assumes a single covariance matrix that applies across clinical contexts. But the cross-dimensional dependencies that \Sigma encodes may vary:

• Across specialties: Surgery and psychiatry may have different \Sigma matrices. Surgery weights d_1 (outcomes) and d_2 (non-maleficence) heavily; psychiatry weights d_4 (autonomy) and d_7 (identity) heavily. The off-diagonal terms may differ correspondingly.

• Across cultures: Japanese medicine, with its emphasis on family-centered decision-making, likely has different \Sigma_{46} (autonomy \times social impact) than American medicine. Middle Eastern medical cultures, with strong emphasis on religious identity, likely have different \Sigma_{27} (rights \times identity). Is there a universal clinical metric, or is the metric itself culturally determined?

• Across patient populations: The covariance matrix as experienced by a disabled patient may differ from the matrix as experienced by a non-disabled patient — not because the manifold structure is different but because the dimensional weights are patient-specific. This raises the question of whether \Sigma should be estimated at the population level (one matrix per culture/specialty) or at the individual level (one matrix per patient).

Question 2: Can \Sigma be estimated from EHR data without prospective instruments?

The experimental designs in Chapter 17 require dedicated data collection — ethics consultation coding, discrete choice experiments, proxy instrument administration. Is it possible to estimate \Sigma from routine EHR data alone, without prospective data collection?

The challenge is that three of nine dimensions (d_5: trust, d_7: dignity, d_9: epistemic status) are not routinely captured in EHRs. Natural language processing of clinical notes may provide partial proxies — mentions of “patient expresses concern about treatment,” “family meeting conducted,” “patient does not understand diagnosis” — but these are noisy and incomplete.

A promising approach: use the six dimensions that are measured in EHRs to estimate a 6 \times 6 submatrix of \Sigma, then use proxy instruments on a subsample to estimate the remaining 27 entries and validate the NLP-based proxies against the instrument-based measurements.

Question 3: How stable is \Sigma over time?

Clinical practice evolves. Patient expectations change. Institutional norms shift. The covariance matrix estimated from 2020 ethics consultations may not apply to 2030 clinical decisions. Is \Sigma stable enough over time to support longitudinal policy evaluation, or must it be re-estimated periodically?

The Noether Connection

If the clinical Lagrangian is invariant under the medical BIP — clinical evaluations do not change when the patient is re-described in equivalent terms — then Noether’s theorem implies a conserved quantity. What is it?

In Geometric Ethics, the BIP-invariance of the moral Lagrangian implies conservation of harm: harm cannot be created or destroyed by re-description, only by real transformation. The clinical analogue should be a conservation law for some clinical quantity — but which one?

Candidate Conservation Laws

Candidate 1: Conservation of clinical-moral status. The patient’s nine-dimensional attribute vector \mathbf{a}(v) is preserved under re-description. The same patient described by two different physicians, in two different languages, using two different chart formats, should have the same attribute vector. This is definitional — it restates the BIP — and therefore does not constitute an independent conservation law.

Candidate 2: Conservation of manifold cost. The total clinical friction along any path is invariant under patient re-description. This is a stronger claim: not just that the patient’s state is invariant but that the cost of the path through the manifold is invariant. If true, it implies that the cost of treating a patient does not depend on how the patient is described — a testable claim that connects to the algorithmic bias literature (where costs demonstrably vary with description).

Candidate 3: Conservation of boundary penalty. The total boundary penalty incurred along a clinical path is invariant under re-description. This is the clinical analogue of harm conservation: the total moral cost of boundary crossings cannot be changed by re-describing the patient.

Candidate 4: No conservation law exists. The clinical Lagrangian may not be exactly BIP-invariant — there may be genuine BIP violations (cases where re-description legitimately changes the evaluation) that break the symmetry and prevent a conservation law from holding. Cultural variation in \Sigma (Question 1 above) would imply that the Lagrangian is not universally invariant, and Noether’s theorem would not apply globally.

The question is open. Identifying the clinical conservation law — or proving that none exists — is a major theoretical challenge for the geometric programme.

Beyond Candidates: A Research Program

The conservation law question (Section 18.2) is not merely academic. If a conservation law for clinical reasoning exists, it would have profound practical implications:

1. Diagnostic power. A conservation law would imply that certain quantities are preserved across clinical re-descriptions. If the patient’s total manifold position is conserved under re-description, then any clinical evaluation that violates conservation — that “creates” or “destroys” manifold position through re-description — is identified as a gauge violation.

2. Constraint on clinical AI. A conservation law would constrain what clinical AI systems can and cannot do. Just as conservation of energy constrains physical systems (you cannot build a perpetual motion machine), conservation of clinical-moral status would constrain clinical systems (you cannot improve a patient’s manifold position merely by re-describing them).

3. Measurement tool. Violations of conservation laws are empirically detectable. If a conservation law holds, its violations are markers of measurement error, gauge violation, or genuine clinical change — and distinguishing among these three is diagnostically valuable.

The research program to derive clinical conservation laws would proceed in three stages:

Stage 1: Formalize the clinical Lagrangian. The Lagrangian \mathcal{L} for the clinical decision complex must be specified — what quantity is extremized along the clinical geodesic? The natural candidate is clinical friction \text{CF}(\gamma), the total cost of the clinical path. The clinical geodesic minimizes \text{CF}, so \text{CF} plays the role of the action in classical mechanics.

Stage 2: Identify the symmetries. The medical BIP states that the Lagrangian is invariant under patient re-description. This is a continuous symmetry group — the group of meaning-preserving transformations of clinical information. The group’s generators must be identified explicitly.

Stage 3: Apply Noether’s theorem. For each continuous symmetry of the Lagrangian, Noether’s theorem yields a conserved quantity. The conserved quantities are the conservation laws of clinical reasoning.

This program has not been completed. The clinical Lagrangian has been informally specified (as clinical friction) but not formally constructed with the variational machinery needed for Noether’s theorem. The symmetry group has been identified in principle (the medical BIP) but its generators have not been enumerated. The program is feasible — it requires the tools of variational calculus on simplicial complexes — but the mathematical development remains to be done.

The Riemannian Assumption

The framework assumes that the clinical decision complex has the structure of a smooth Riemannian manifold with a positive-definite metric. This assumption enables the use of geodesics, curvature, and parallel transport. But clinical decision-making may have geometric structures that Riemannian geometry cannot capture.

Three Non-Riemannian Possibilities

Possibility 1: Singularities.

Clinical decision-making may have singularities — points where the metric becomes degenerate (the determinant of \Sigma approaches zero). A natural candidate: diagnostic uncertainty. When the diagnosis is unknown, the patient’s position on the clinical manifold is indeterminate — the metric in the diagnostic dimension becomes degenerate because the clinician does not know which region of the manifold the patient occupies.

If the clinical manifold has singularities, the geodesic equation breaks down at those points — the clinician cannot compute the minimum-cost path because the cost function is undefined. This may explain the phenomenon of “diagnostic paralysis” — the clinician’s inability to decide when the diagnosis is uncertain. The paralysis is not cognitive failure; it is a geodesic singularity.

Possibility 2: Non-Hausdorff topology.

In Riemannian geometry, the manifold is Hausdorff — any two distinct points can be separated by open neighborhoods. But clinical diagnosis may violate this: two diseases with overlapping presentations (e.g., pulmonary embolism and pneumonia in an elderly patient with dyspnea) may occupy regions of the manifold that cannot be separated by any diagnostic test. The clinical state is genuinely ambiguous — the patient is “between” two diagnoses in a way that the Hausdorff property does not allow.

Non-Hausdorff manifolds are mathematically exotic, but they may be clinically natural. The framework would need to be extended to accommodate diagnostic indistinguishability as a topological rather than merely epistemic phenomenon.

Possibility 3: Finsler structure.

The Riemannian metric is symmetric: the distance from state A to state B equals the distance from state B to state A. But clinical transitions may be asymmetric: the cost of transitioning from health to illness is not the same as the cost of transitioning from illness to health. Recovery is not the reverse of disease onset — it follows a different path through the manifold, with different costs on different dimensions.

A Finsler metric — a metric that depends not just on position but on direction — would capture this asymmetry. The clinical geodesic from illness to health would differ from the geodesic from health to illness, even between the same two points. This would explain the clinical observation that recovery paths are not reverse-disease paths: a patient who develops heart failure through a specific sequence of events does not recover by reversing that sequence.

Research Direction

Determining whether the clinical manifold is Riemannian, Finsler, or topologically exotic requires empirical data about the symmetry properties of clinical transitions. The discrete choice experiments proposed in Chapter 17 can test for asymmetry: do clinicians assign different costs to the transition A \to B and the transition B \to A? If costs are systematically asymmetric, the Finsler hypothesis is supported.

The General Theorem and Its Clinical Specificity

The No Escape Theorem (Chapter 16) states that an AI system constrained by the clinical decision complex cannot produce recommendations that cross sacred-value boundaries. The general theorem is proved. The clinical-specific questions are:

Question 4: What are the precise sacred-value boundaries for medical AI?

The framework identifies three candidates: non-consensual experimentation (\beta = \infty), deliberate non-palliative killing (\beta = \infty), and patient instrumentalization (\beta = \infty). But the full set of sacred-value boundaries for medical AI has not been enumerated. Are there AI-specific boundaries — constraints that apply to artificial pathfinders but not to human clinicians?

Candidate: the opacity boundary — a constraint that an AI system’s recommendation must be explainable to the patient and the treating clinician. If the AI produces a recommendation that it cannot explain (the “black box” problem), the recommendation crosses a boundary on d_9 (epistemic status) and possibly d_8 (institutional legitimacy) that has infinite cost. An AI system constrained by the opacity boundary cannot produce unexplainable recommendations, regardless of their d_1 optimality.

Question 5: What are the engineering requirements for manifold-constrained AI?

The No Escape Theorem is a mathematical result about the properties of pathfinding on a manifold with infinite-cost boundaries. Translating this result into engineering requirements for actual AI systems is a separate challenge. What does a manifold-constrained medical AI look like in practice?

Possibilities include: - Hard-coded boundary constraints in the AI’s reward function — but these are parametric guardrails, not structural constraints, and can be breached by distribution shift. - Manifold-embedded architectures where the AI’s output space is structurally limited to the manifold’s interior — but this requires encoding the manifold’s geometry into the AI’s architecture, which has not been done. - Proof-carrying recommendations where each AI recommendation comes with a certificate that no sacred-value boundary was crossed — but this requires a formal verification system that does not yet exist for clinical AI.

Question 6: Can an AI system be provably incapable of crossing sacred-value boundaries?

This is the strongest possible safety guarantee: not just that the AI is unlikely to cross boundaries, or that it will be caught if it does, but that crossing is impossible by the architecture of the system. The No Escape Theorem says this is mathematically possible. Whether it is engineerable is an open question in AI safety.

The Decisive Test

The Moral Injury Index \text{MI}(P) is designed to be computed before a policy is implemented, predicting the moral injury that the policy will produce. The decisive validation test is:

1. Compute \text{MI}(P) for a proposed policy.
2. Implement the policy.
3. Measure actual moral injury in the clinician workforce.
4. Compare predicted and actual MI.

If predicted and actual MI correlate (r > 0.5), the Moral Injury Index has predictive validity and can be used as a prospective policy evaluation tool. If they do not correlate, the Index’s mechanism — that moral injury is a function of boundary-crossing frequency and clinician \beta_k profiles — is wrong or inadequately operationalized.

Question 7: Has any institution used the framework to redesign a policy and measured the resulting change in clinician MI?

This is the ultimate applied test. The framework does not merely describe moral injury — it claims to predict it and to guide prevention. An institution that computes \text{MI}(P), redesigns the policy to reduce boundary crossings, and measures a reduction in clinician moral distress would provide the strongest possible evidence for the framework’s practical value.

No such test has been conducted. It requires institutional buy-in (a hospital willing to compute the MI Index, redesign a policy, and measure outcomes), clinical ethics expertise (to code boundary crossings and estimate \beta_k distributions), and longitudinal data collection (moral distress measurements before and after the policy change). The requirements are feasible but demanding.

Nine Dimensions: Sufficient or Not?

The nine dimensions of the clinical decision complex are inherited from the nine-dimensional moral manifold of Geometric Ethics. They were not derived from clinical data — they were derived from the structure of moral reasoning and instantiated for clinical contexts. The question is whether nine dimensions are sufficient for clinical medicine, or whether the clinical domain requires additional dimensions.

Candidate Additional Dimensions

Time pressure. Clinical decisions made under time pressure (emergency triage) differ from decisions made at leisure (elective surgery planning). The current framework captures time pressure indirectly — through its effect on heuristic quality (time pressure degrades h(n)) and through the distinction between acute and chronic settings. But time pressure may be an independent dimension that affects the clinical manifold’s structure in ways that the current nine dimensions do not capture.

Resource scarcity. The availability of resources (ventilators, ICU beds, medications) constrains the set of feasible paths on the manifold. The current framework models scarcity as a constraint on the triage optimization (\sum_i r(\gamma_i) \leq R). But scarcity may also affect the edge weights — the moral cost of a treatment is higher when the treatment consumes a scarce resource, even for an individual patient outside the triage context.

Team dynamics. Clinical decisions are rarely made by a single clinician. The interaction between team members — attending, resident, nurse, social worker, chaplain — introduces a multi-agent dimension that the current framework models as a single decision-maker with a composite heuristic. The team dynamics may be an independent dimension, especially in settings where team conflict affects clinical decisions.

Question 8: Are some dimensions redundant in specific clinical contexts?

Conversely, some of the nine dimensions may be redundant — highly correlated or functionally identical — in specific clinical contexts. In routine outpatient care (e.g., managing a patient’s hypertension), dimensions d_3 (justice), d_6 (social impact), and d_8 (institutional legitimacy) may be minimally activated, reducing the effective dimensionality to six or fewer. If the manifold’s effective dimensionality varies by context, the framework should acknowledge this variation — and the measurement strategy should accommodate it.

The Empirical Mapping Challenge

The boundary penalties \beta_k and the covariance matrix \Sigma are not universal constants. They vary across clinical cultures:

• Autonomy-centered cultures (United States, Northern Europe): High \beta_{\text{consent}}, high weight on d_4. Patient autonomy is the dominant value; paternalism is strongly penalized.

• Family-centered cultures (Japan, much of East Asia, Mediterranean cultures): Lower \beta_{\text{consent}} for family-mediated consent, high weight on d_6 (social and family impact). The family is the unit of decision-making, not the individual.

• Community-centered cultures (many Indigenous traditions, some African and South Asian cultures): High weight on d_6 and d_3 (justice at the community level). The community’s needs may override individual autonomy in ways that autonomy-centered cultures would find unacceptable.

• Religious cultures (many Middle Eastern and South Asian contexts): High weight on d_7 (identity as tied to religious practice) with specific boundary penalties for interventions that conflict with religious obligations.

Question 9: Can the framework accommodate cross-cultural variation without relativism?

The framework must navigate between two extremes: universalism (one set of \beta_k values for all cultures — which ignores genuine cultural variation) and relativism (every culture’s \beta_k values are equally valid — which provides no basis for identifying unjust practices within any culture).

The geometric framework offers a middle path: the manifold structure is universal (nine dimensions, the BIP, the pathfinding model), but the metric is culturally variable (\Sigma and \beta_k are context-dependent). This means that the framework can accommodate cultural variation — different cultures assign different costs to different boundary crossings — while maintaining the structural apparatus that enables cross-cultural comparison (the Bond Index, the gauge-invariance test, the Moral Injury Index).

The cultural variation must be mapped empirically. This requires estimating \Sigma in multiple cultural contexts and identifying which entries are culturally invariant (if any) and which are culturally determined.