Chapter 7: Clinical Heuristics and the \varepsilon-Admissibility Spectrum

7.1 Admissibility: The Central Concept

In A* search, a heuristic h(n) is admissible if it never overestimates the true remaining cost:

h(n) \leq h^*(n) \quad \text{for all } n

where h^*(n) is the true optimal cost from n to the nearest goal in G. When the heuristic is admissible, A* is guaranteed to find the optimal path — the clinical geodesic. When the heuristic is inadmissible (it overestimates the cost somewhere), A* may miss the optimal path by avoiding regions of the search space that the heuristic incorrectly labels as expensive.

The clinical instantiation is immediate:

Theorem 7.1 (Clinical Heuristic Admissibility). Let h_C(n) be a clinical moral heuristic — a function that assigns high cost to actions violating professional ethical norms and low cost to actions consistent with them:

h_C(n) = \sum_k \beta_k \cdot P(\text{remaining path from } n \text{ crosses boundary } k)

If the boundary penalties \beta_k are calibrated by clinical training and professional socialization to be lower bounds on the true moral cost of boundary crossing (i.e., \beta_k \leq \beta_k^*, where \beta_k^* includes the full cost: malpractice liability, professional censure, patient harm, institutional trust damage, and psychological distress to the clinician), then h_C is admissible and A* search finds the optimal clinical path.

Proof. The clinical heuristic h_C(n) = \sum_k \beta_k \cdot P_k(n) estimates the remaining cost as the expected boundary penalty cost. Each \beta_k \leq \beta_k^* by assumption, and the remaining Mahalanobis distance cost (which h_C ignores) is non-negative. Therefore h_C(n) \leq h^*(n), satisfying the admissibility condition. By the correctness of A* with admissible heuristics (Hart, Nilsson, & Raphael, 1968), the search finds the minimum-cost path. \square

The theorem’s clinical content is this: if medical training calibrates moral intuitions as lower bounds on the true cost of ethical violations — if the clinician learns to feel at least as much discomfort from boundary crossings as the full institutional, legal, relational, and psychological cost would justify — then the clinician’s pathfinding will find the optimal clinical path. The heuristic will guide the search toward the geodesic, not away from it.

The question is: what happens when the calibration fails?

7.2 The Four-Category Hierarchy

Clinical heuristics form a hierarchy of four categories, ordered by their relationship to admissibility. Each category has a distinctive clinical signature and a distinctive failure mode.

7.3 The Anatomy of \varepsilon

The error \varepsilon in the \varepsilon-admissibility condition is not a single number; it is dimension-specific. A clinician may have excellent calibration on d_1 (\varepsilon_1 \approx 0.05) and poor calibration on d_7 (\varepsilon_7 \approx 0.5). The overall \varepsilon is the maximum dimension-specific error:

\varepsilon = \max_k \varepsilon_k

This decomposition is clinically useful. It identifies where the clinician’s heuristic is most miscalibrated and therefore where training or recalibration efforts should be focused.

Proposition 7.2 (Dimension-Specific \varepsilon-Profiles). Different clinical specialties have different \varepsilon-profiles:

• Emergency medicine: Low \varepsilon_1 (excellent d_1 calibration under acute conditions), high \varepsilon_7 (poor d_7 calibration due to time pressure that prevents dignity considerations).
• Palliative care: Low \varepsilon_7 (excellent d_7 calibration), moderate \varepsilon_1 (outcome calibration is less critical when outcomes are secondary to process).
• Surgery: Low \varepsilon_1 (excellent outcome calibration), high \varepsilon_4 (autonomy calibration may be poor due to the paternalistic tradition of surgical culture).
• Psychiatry: Low \varepsilon_4 (excellent autonomy calibration, given the constant engagement with consent and capacity issues), high \varepsilon_8 (institutional calibration may be poor due to tension between clinical judgment and institutional mandates).

These profiles are testable predictions. The framework predicts that specialty-specific heuristic miscalibrations will manifest as specialty-specific patterns of clinical decision suboptimality — and that cross-specialty ethics consultations are effective precisely because they bring complementary \varepsilon-profiles to a shared clinical problem.

7.4 Clinical Training as Heuristic Calibration

The central thesis of this section is that the primary function of clinical training — medical school, residency, fellowship, continuing medical education — is not to teach clinicians to compute g(n) (evidence-based medicine is taught through didactics and self-study) but to calibrate h(n) (clinical wisdom is learned through supervised pathfinding on the clinical decision complex).

7.5 Ethics Consultations as Heuristic Recalibration

When a clinician requests an ethics consultation, the geometric interpretation is: “My h(n) is producing a path that I am uncomfortable with, and I need help recalibrating.”

The ethics consultant brings three resources to the recalibration:

1. A different h(n). The consultant has different experiences, different boundary penalties, and different dimensional weightings. The consultant’s h(n) may identify boundaries that the primary clinician has not noticed, or assign different costs to boundaries that the primary clinician has noticed but weighted differently.

2. An explicit dimensional decomposition. The primary clinician’s h(n) operates automatically, without conscious access to the dimensional weights. The ethics consultant makes the weights explicit: “You’re weighing autonomy heavily here, but have you considered the trust implications?” This explicitness allows the clinician to examine their heuristic and adjust it consciously.

3. Population-level calibration data. The ethics consultant has seen many similar cases and has calibrated their heuristic across a population. This population-level calibration can correct the primary clinician’s case-specific calibration, which may be skewed by the current patient’s particular features.

Proposition 7.3 (Ethics Consultations Reduce \varepsilon-Variance). The framework predicts that effective ethics consultations should reduce the variance of \beta_k estimates in the clinician population (making heuristics more consistent across clinicians) without reducing the mean (maintaining appropriate boundary sensitivity). Consultations that reduce both variance and mean are not recalibration; they are desensitization — a form of institutional moral numbing.

This prediction is testable. Survey clinicians before and after ethics consultations for their \beta_k estimates on the case’s relevant boundaries. The prediction is that the post-consultation variance is lower (clinicians converge toward agreement) while the mean is stable (the group’s average boundary sensitivity is preserved).

7.6 Defensive Medicine as Heuristic Inflation

Defensive medicine — ordering tests, treatments, or procedures to reduce legal risk rather than to improve clinical care — is the most common form of institutional heuristic inflation in contemporary medicine.

The geometric analysis is precise. Defensive medicine inflates \beta_{\text{liability}}: the boundary penalty for “failing to test” or “failing to treat” is set higher than the true clinical-moral cost of the failure, because the legal cost of a missed diagnosis is added to the moral cost. The inflation is rational from the individual clinician’s perspective: the legal system imposes financial and reputational costs that exceed the moral cost of the diagnostic error. But the inflation distorts the clinical geodesic: the path that minimizes legal risk is not the path that minimizes total clinical-moral friction.

Proposition 7.4 (Defensive Medicine as Geodesic Distortion). Let \gamma_C^* be the clinical geodesic computed with true boundary penalties \beta_k^*, and let \gamma_D^* be the geodesic computed with liability-inflated boundary penalties \beta_k^D > \beta_k^*. Then \gamma_D^* systematically over-tests (higher d_3 cost from resource consumption and higher d_1 cost from iatrogenic harm of unnecessary procedures), over-treats (higher d_7 cost from interventions the patient does not need or want), and under-communicates (higher d_9 cost from spending time on documentation rather than patient communication).

The moral injury implications are significant. When the institution’s legal environment inflates \beta_{\text{liability}} beyond the clinician’s moral calibration, the clinician is forced to follow a path (\gamma_D^*) that their own h(n) identifies as suboptimal. This is a forced boundary crossing of a different kind: not the crossing of a moral boundary but the avoidance of the geodesic under institutional pressure. The clinician knows that the extra test is unnecessary, that the additional treatment is not indicated, that the documentation time would be better spent talking to the patient. The institutional geodesic mandates the detour anyway. The resulting moral distress is the heuristic’s protest: the clinician’s h(n) is correctly calibrated, but the institutional \beta_{\text{liability}} is overriding it.

7.7 Sarah’s Heuristic Calibration

Sarah’s heuristic function at the start of residency was uncalibrated. She had two specific miscalibrations that Dr. Torres identified and corrected:

Inflated \beta_{\text{difficult conversation}}. Sarah avoided goals-of-care conversations because she overestimated their cost. She imagined that the conversation would be painful (d_7: the patient would be distressed), damaging to trust (d_5: the patient would lose confidence), and futile (d_1: the conversation would not change the clinical outcome). Her \beta for “having the difficult conversation” was approximately 15 — higher than her \beta for “doing nothing and hoping the situation resolves.”

Dr. Torres corrected this by supervised pathfinding. She took Sarah into goals-of-care conversations, showed her how to navigate them, and demonstrated that the conversations typically improved d_5 (patients trust clinicians who are honest), improved d_7 (patients feel dignified when their values are heard), and improved d_9 (patients understand their situation better after the conversation). Sarah’s \beta_{\text{difficult conversation}} dropped from 15 to 3 over the course of six months of supervised practice.

Deflated \beta_{\text{watchful waiting}}. Sarah underestimated the cost of delay. She was comfortable with “let’s see how this develops” because it avoided the immediate cost of intervention. Her \beta for “waiting” was approximately 1 — lower than her \beta for “acting,” even when the waiting allowed the patient to deteriorate.

Dr. Torres corrected this by showing Sarah the trajectory costs of delay: the patient whose sepsis was not treated for four hours because “let’s see if the fever resolves,” the patient whose cancer diagnosis was delayed by three months because “let’s repeat the imaging in six months.” Sarah’s \beta_{\text{watchful waiting}} increased from 1 to 5 over the course of the year, reflecting a more accurate estimate of the cost of delay on d_1 and d_9.

By the end of her second year, Sarah’s h(n) was \varepsilon-admissible with \varepsilon \approx 0.15 — still far from the \varepsilon \approx 0.05 of an experienced attending, but well within the range of competent clinical reasoning. She could navigate the manifold with reasonable accuracy, recognizing boundaries before she reached them, estimating remaining costs within a factor of 1.15 of the true costs, and seeking help (ethics consultation, attending supervision) when she recognized that her heuristic was uncertain.

Three years later, the pandemic would damage her heuristic. Chapters 13–15 describe what happened to her h(n) when she was forced to cross boundaries she had been carefully trained to respect.

7.8 The Interaction Between \varepsilon-Admissibility and Moral Injury

The relationship between the heuristic hierarchy and moral injury (Chapters 13–15) is bidirectional and clinically consequential.

Moral injury degrades admissibility. A clinician who begins their career with an \varepsilon-admissible heuristic (\varepsilon \approx 0.1) may, after repeated forced boundary crossings, develop an inadmissible heuristic on the violated dimensions. The inflation or deflation of boundary penalties shifts the heuristic out of the \varepsilon-admissible range on specific dimensions while leaving it calibrated on others. The result is a clinician who makes excellent decisions in one domain (e.g., diagnostic reasoning, where d_1 and d_9 are primary) but systematically suboptimal decisions in another domain (e.g., triage, where the violated boundary is active).

Proposition 7.5 (Moral Injury as Dimensional \varepsilon-Degradation). Moral injury produces dimension-specific \varepsilon-degradation: the overall heuristic error \varepsilon = \max_k \varepsilon_k increases not because all dimensions are miscalibrated but because the violated dimension’s \varepsilon_k has increased substantially. The clinical signature is domain-specific avoidance or desensitization that does not affect unrelated clinical domains.

This prediction is testable. Survey clinicians who report moral injury for their \beta_k estimates on the violated boundary type. Compare to clinicians without moral injury on the same boundary type. The prediction is that morally injured clinicians show significantly higher \varepsilon_k (either inflated \beta_k for hypervigilant clinicians or deflated \beta_k for morally numb clinicians) on the specific boundary where the injury occurred, with no significant difference on other boundaries.

Admissibility analysis identifies the injury mechanism. Conversely, the heuristic hierarchy provides a diagnostic tool for moral injury: by measuring the dimension-specific \varepsilon_k profile of a clinician, one can identify which boundaries have been damaged and in which direction (inflation vs. deflation). This moves moral injury from a global assessment (“this clinician is distressed”) to a dimensional assessment (“this clinician has inflated \beta_{\text{futile}} and deflated \beta_{\text{dignity}} as a result of palliative care moral injury”). The dimensional assessment points to a specific recalibration strategy: restore \beta_{\text{futile}} to its appropriate value through structured case review, and restore \beta_{\text{dignity}} through supervised dignity-preserving care.

7.9 Institutional Heuristics

The chapter has focused on individual clinician heuristics. But institutions also have heuristics — institutional norms, protocols, guidelines, and cultural expectations that function as collective h(n) estimates. Institutional heuristics interact with individual heuristics and can either support or undermine individual calibration.

Supportive institutional heuristics. A hospital culture that values goals-of-care conversations, provides protected time for consent processes, supports ethics consultations, and structures M&M conferences for constructive heuristic recalibration provides an institutional environment that supports individual \varepsilon-admissibility. The institution’s collective h(n) — its norms about what matters and how to navigate the manifold — is aligned with the individual clinician’s heuristic, creating a feedback loop that maintains calibration.

Undermining institutional heuristics. A hospital culture that values throughput over process, penalizes clinicians for taking time with patients, discourages ethics consultations as “slowing things down,” and uses M&M conferences for blame rather than learning provides an institutional environment that systematically degrades individual \varepsilon-admissibility. The institution’s collective h(n) — which emphasizes d_1 and d_8 at the expense of d_4, d_5, d_7, and d_9 — forces individual clinicians to override their own calibrated heuristics in favor of the institutional heuristic, producing moral distress and eventual moral injury.

Proposition 7.6 (Institutional Heuristic Determines Individual \varepsilon-Trajectory). A clinician’s long-term \varepsilon-trajectory — whether their heuristic improves, stagnates, or degrades over time — is determined more by the institutional heuristic environment than by the clinician’s initial calibration or individual training. Well-calibrated clinicians in undermining institutional environments degrade toward inadmissibility. Poorly calibrated clinicians in supportive institutional environments improve toward \varepsilon-admissibility. The institution is the clinician’s heuristic ecosystem.

This proposition has direct implications for physician retention and workforce health. It predicts that clinician moral distress and attrition correlate not with the inherent difficulty of the clinical work (which determines the frequency of boundary encounters) but with the institutional heuristic environment (which determines whether boundary encounters degrade or maintain the clinician’s heuristic). A difficult job in a supportive environment produces calibrated clinicians. An easy job in an undermining environment produces damaged clinicians.

7.10 Summary Table: The Four Categories

The hierarchy is the diagnostic framework for clinical judgment quality. By identifying which category a clinician’s heuristic falls into — and on which dimensions — the framework enables targeted interventions that restore the heuristic toward \varepsilon-admissibility. The interventions are not punitive; they are calibrative. A clinician with an inadmissible heuristic is not a bad clinician; they are a damaged clinician whose h(n) has been altered by forces beyond their control. The institutional response should be recalibration, not discipline.

Summary

This chapter has characterized the full spectrum of clinical heuristic quality through a four-category hierarchy: strictly admissible (prohibitions whose true cost exceeds any estimate), \varepsilon-admissible (well-trained clinicians whose boundary penalties are approximately correct), inadmissible (clinicians with inflated boundary penalties from moral injury, defensive medicine, or personal trauma), and gauge-variant (clinicians whose heuristics depend on framing rather than clinical reality). Medical training is heuristic calibration — the reduction of \varepsilon from the large values of untrained clinicians to the small values of experienced practitioners. Ethics consultations are heuristic recalibration — the correction of inadmissible heuristics toward \varepsilon-admissibility. Defensive medicine is heuristic inflation — the systematic overestimation of boundary penalties due to institutional pressure.

The heuristic hierarchy generates testable predictions: specialty-specific \varepsilon-profiles (emergency medicine has low \varepsilon_1 and high \varepsilon_7), the effectiveness of ethics consultations (variance reduction without mean shift), the clinical signature of moral injury (systematic avoidance of specific boundary crossings, measurable as inflated \beta_k values on specific dimensions), and the role of institutional culture in determining long-term heuristic trajectories.