curiositech/ndm-law-enforcement

Naturalistic Decision Making for High-Stakes Domains

DECISION POINTS

Primary Decision Tree: Recognize-and-Act vs Deliberate-and-Compare

Is the situation pattern-matchable to prior experience?
├── YES: Generate one plausible action
│   ├── Mental simulation passes → ACT immediately
│   └── Simulation fails → Modify action OR generate next best candidate
└── NO: Is this genuine novelty or pattern library failure?
    ├── Genuine novelty → Deliberate carefully, document for future patterns
    └── Pattern gap → Flag training need, proceed with explicit uncertainty

Time Pressure Assessment

Time available vs. cost of error:
├── High time pressure + Low error cost → Act on current best read
├── High time pressure + High error cost → Quick situation assessment, then act
├── Low time pressure + Any error cost → Full situation assessment required
└── Conflicting cues present → ALWAYS invest in situation clarification first

Expertise Level Calibration

Agent expertise level assessment:
├── Novice/Advanced Beginner → Strict rule following, no improvisation
├── Competent → Rules apply, exceptions require explicit justification (DANGER ZONE)
└── Proficient/Expert → Pattern recognition reliable, intuition is valid data

Post-Decision Failure Diagnosis

Decision failed - where was the breakdown?
├── Wrong situation read → Failure in SITUATION ASSESSMENT
│   └── Review cue interpretation, not action selection
├── Right read, wrong action → Failure in ACTION SELECTION
│   └── Review option generation or mental simulation
├── Acted before understanding → PREMATURE CLOSURE
│   └── Novice/Competent behavior pattern
└── Analyzed past action window → DECISION PARALYSIS
    └── Missing satisficing rules

FAILURE MODES

1. Tunnel Vision Syndrome

Symptom: Agent fixates on initial interpretation despite contradicting evidence Root Cause: Premature closure - commitment to first plausible story without verification Detection Rule: If agent ignores 3+ cues that contradict primary hypothesis, tunnel vision active Fix: Force hypothesis generation of 2+ alternative explanations before action

2. Option Enumeration Paralysis

Symptom: Agent generates exhaustive option lists under time pressure Root Cause: Novice pattern mistaken for rigor - treating optimization as always superior to satisficing Detection Rule: If agent spends >30% of available time on option generation vs. situation assessment, paralysis mode active Fix: Implement explicit satisficing thresholds and stop-search rules

3. Competent Stage Overconfidence

Symptom: Agent breaks rules without situational justification Root Cause: Enough experience to see exceptions, insufficient models to know when rule-breaking is safe Detection Rule: If agent deviates from protocols but cannot articulate situation-specific reasoning, overconfidence active Fix: Require explicit situational model before any rule deviation

4. Action-First Assessment

Symptom: Agent jumps to response selection before understanding what's happening Root Cause: Confusing speed of action with speed of assessment Detection Rule: If agent selects response before completing "what/why/what-next" situation model, action-first active Fix: Mandatory situation assessment checkpoint before action consideration

5. Expertise Misattribution

Symptom: System fails because expert self-reports were taken at face value Root Cause: Assuming experts can accurately describe their own decision processes Detection Rule: If knowledge base built from direct interviews without structured elicitation, misattribution likely Fix: Use CDM (Critical Decision Method) structured retrospective interviews

WORKED EXAMPLES

Example 1: Ambush Scenario - Expert vs Novice Decision Trees

Setup: Officer approaches vehicle during traffic stop. Driver's hands suddenly drop below window line.

Novice Decision Tree:

1. Notices hand movement (single cue)
2. Recalls training rule: "Hands not visible = potential threat"
3. Draws weapon immediately
4. Miss: No assessment of totality - time of day, passenger behavior, driver's verbal responses
5. Outcome: Escalation of non-threatening situation

Expert Decision Tree:

1. Notices hand movement (primary cue)
2. Situation Assessment Phase:
   • Passenger behavior (calm vs. agitated)
   • Driver verbal responses (compliant vs. evasive)
   • Environmental context (lighting, escape routes, backup proximity)
   • Vehicle contents visible
3. Pattern Recognition: "Nervous compliance" vs. "Pre-attack indicators"
4. Mental Simulation: "If I draw now, what happens next?"
5. Action: Verbal command + tactical positioning (vs. immediate weapon draw)
6. Continuous Assessment: Ready to escalate but hasn't committed to single interpretation

Key Difference: Expert invests cognitive resources in situation model before action selection. Novice applies rule to single cue.

Example 2: Hostage Situation - Satisficing vs. Optimization

Setup: Armed subject holding hostage, demands negotiation. SWAT has clear shot opportunity.

Optimization Approach (Failure Mode):

• Enumerate all possible interventions
• Analyze probability matrices for each
• Seek additional intelligence to improve analysis
• Time Cost: 45+ minutes of analysis
• Outcome: Subject escalates during delay, situation deteriorates

Expert Satisficing Approach:

• Quick Assessment: Subject behavior pattern (calm/agitated), hostage condition, containment status
• First Viable Option: Negotiation attempt with sniper backup positioned
• Stop-Search Rule: "If negotiation maintains status quo for 10 minutes and subject shows compliance indicators, continue. If escalation indicators appear, transition to tactical solution"
• Mental Simulation: "This approach handles the 80% probability scenarios adequately"
• Time Cost: 5 minutes to implementation
• Outcome: Faster stabilization, preserved options

Trade-off Recognition: Expert accepts "good enough" solution quickly rather than pursuing optimal solution slowly. The risk of delay exceeded the benefit of analysis perfection.

QUALITY GATES

Situation assessment is complete when:

• [ ] Primary story can be articulated: "What is happening here?"
• [ ] Key cues supporting story have been identified (minimum 3)
• [ ] Alternative explanations considered: "What else could this be?"
• [ ] Prediction generated: "What happens next if story is correct?"
• [ ] Contradicting evidence actively sought: "What would prove me wrong?"
• [ ] Time/stakes calibration: "How much certainty does this situation require?"
• [ ] Expert intuition acknowledged: "Does something feel off?"

Action selection is complete when:

• [ ] Response matches situation assessment (not generic protocol)
• [ ] Mental simulation of consequences performed
• [ ] Fallback options identified if primary response fails

Decision implementation is complete when:

• [ ] Monitoring plan active for disconfirming cues
• [ ] Escalation/de-escalation triggers defined
• [ ] Success/failure criteria specified

NOT-FOR Boundaries

Do NOT use this skill for:

• Routine procedural decisions → Use standard operating procedures instead
• Technical/engineering problems → Use systematic analysis frameworks instead
• Novel situations with unlimited time → Use deliberate problem-solving methods instead
• Decisions where failure cost is catastrophic → Use formal risk analysis instead

When to delegate:

• For pure optimization problems → Use operations research methods
• For complex system design → Use systems engineering approaches
• For policy development → Use structured policy analysis frameworks
• For training design → Use instructional design methodologies

This skill is specifically for:

• Time-pressured decisions with incomplete information
• Replicating expert judgment under uncertainty
• Diagnosing decision failures in dynamic environments
• Building agent architectures for high-stakes domains