Expert Decision-Making Is Not Intuition—It’s an Observable, Verifiable Behavioral Contract

Rule One: Recalculate probabilities only when new evidence objectively shifts judgment. Trigger conditions are strictly observable: ① Evidence comes from independent sources (e.g., two physicians from different specialties, or two physically isolated sensors); ② The statistical divergence (measured via KL divergence) between the updated and prior belief distributions exceeds a field-calibrated threshold—e.g., 0.25 in emergency triage, validated across 500 historical cases to reduce misclassification by 17% (p<0.01); ③ Full audit trail is generated: prior/posterior distributions, evidence weights, and calculation timestamp. It ignores ‘I sense a change’—and responds only to ‘two independent sources converge on the same anomaly, with statistically significant deviation.’

Rule Two: Every conclusion must be ‘unpackable and inspectable.’ Before output, the human or system must produce structured reasoning: top-level claim (e.g., ‘Proceed immediately to endotracheal intubation’), with mutually exclusive and collectively exhaustive (MECE) supporting arguments—and each intermediate node tied to an operationally verifiable definition. For example, ‘respiratory failure’ isn’t just a term; it must read: ‘SpO₂ < 88% for 60 consecutive seconds, unimproved by supplemental oxygen.’ This structure is saved in a standardized format (e.g., JSON Schema), enabling supervisors to perform a logic-integrity spot-check in under 30 seconds.

Rule Three: When data ‘suddenly stops making sense,’ launch counterfactual simulation. Trigger requires dual observables: ① A key metric (e.g., heart rate variability, order fill latency) breaches its rolling 90-day mean ±3 standard deviations; ② Simultaneously, the current diagnostic model’s prediction residuals exceed ±2σ for three consecutive samples—and with consistent sign (e.g., all three overestimates). The system then auto-generates ≤2 minimal-change hypotheses (e.g., adding only ‘patient took beta-blocker 90 minutes ago’) and projects outcomes to the next measurable node (e.g., ‘expected heart rate drops to 52 bpm within 5 minutes’), broadcasting results live to all collaborators.

Rule Four: When cognitive load peaks, externalize thinking—immediately. All triggers are objectively measurable: ① The same constraint type is missed twice consecutively (e.g., ignoring drug interaction alerts twice); ② Decision nesting exceeds 4 layers (measured by EHR subtask expansion depth or flowchart parenthesis levels); ③ Pressing the ‘intuitive judgment’ button forces immediate generation of a structured hypothesis tree (per Rule Two) with confidence intervals (e.g., ‘70% confident—based on 9 successes in last 12 similar cases’). The entire process is screen-recorded and archived as the sole record for retrospective review.

Rule Five: Handover and pause are driven—not by subjective fatigue—but by objective signals. Any one condition triggers action: ① Samn-Perelli Fatigue Scale score ≥6 (administered via QR-code scan; option text explicitly states: ‘trembling hands, repeated input errors, noticeable reaction delay’); ② Two consecutive critical judgments contain logical contradictions (e.g., ‘BP rise → administer antihypertensive’ and ‘BP rise → indicates hypovolemia → give fluid’ appearing on the same treatment sheet), automatically flagged by an embedded logic validator; ③ Wearable device detects entry into the trough phase of the 90±15-minute ultradian rhythm (via axillary temperature sampled every 15 minutes + moving average), concurrent with self-reported fatigue ≥6. Every trigger auto-generates a handover checklist (pending items, testable assumptions, next decision window) pushed to the relieving colleague’s terminal.