curiositech/law-enforcement-decision-making-during-c

SKILL.md: Naturalistic Decision Making for High-Stakes Domains

license: Apache-2.0

metadata:
  name: naturalistic-decision-making
  version: 1.0
  source: "Zimmerman, L.A. (2006). Law Enforcement Decision Making During Critical Incidents"
  domain: expert-cognition, decision-making, agent-architecture
  activation_triggers:
    - "how should I handle this under time pressure"
    - "expert vs novice decision making"
    - "when to act vs when to analyze"
    - "building knowledge bases from expert knowledge"
    - "why did this decision fail"
    - "how do I handle incomplete information"
    - "pattern recognition vs deliberation"
    - "satisficing vs optimizing"
    - "situation assessment"
    - "tacit knowledge elicitation"

---

When to Use This Skill

Load this skill when facing problems that involve:

• Time-pressured decisions where analytic comparison of all options is not feasible
• Ambiguous situations where the facts are incomplete and action is still required
• Diagnosing decision failures — understanding why something went wrong, not just what went wrong
• Designing agent decision architectures — especially where the agent must behave reliably under uncertainty
• Extracting expert knowledge to encode into systems, prompts, or documentation
• Assessing expertise level of a system or agent to calibrate appropriate behavior
• Distinguishing situation assessment from action selection when debugging poor outcomes

This is not a general optimization or planning skill. It is specifically for understanding and replicating expert cognition in high-stakes, dynamic, time-pressured conditions.

---

Core Mental Models

1. Recognition Before Deliberation

Experts don't compare options — they recognize situations. When a pattern matches a stored mental model, they generate one plausible course of action and mentally simulate its consequences. Only when simulation fails do they generate an alternative. This is faster than analysis and surprisingly reliable — but only if the mental model library is rich enough.

Agent implication: Build pattern-matching first. Reserve option-generation for genuine novelty. Never enumerate all options as a default.

---

2. Situation Assessment Is the Core Skill

The primary separator between expert and novice is not better action selection — it is better understanding of what is happening. Experts invest cognitive resources in building a coherent "situation model" (cues → story → prediction) before selecting action. Novices skip to action scripts.

Agent implication: Diagnosis quality determines outcome quality. A correct action applied to a misread situation fails. Invest in reading the situation before choosing the action.

---

3. The Five-Stage Expertise Arc Has Predictable Failure Modes

Novice → Advanced Beginner → Competent → Proficient → Expert. Each stage has characteristic decision signatures and failure modes. Competent is paradoxically most dangerous: the agent starts breaking rules but lacks the situational models to know when rule-breaking is safe.

Agent implication: Know your level. Novice agents should follow rules strictly. Expert agents can adapt — but only if they have the situational models to justify adaptation. Competent agents claiming expert flexibility are a hazard.

---

4. Satisficing Is Rational, Not a Fallback

Under real-world constraints, the optimal strategy is the first solution that clears an acceptable threshold, not the globally best solution. The goal is to stop searching when a solution is good enough, given: time available, cost of error, and situation familiarity.

Agent implication: Build explicit stop-search rules. Unbounded optimization in time-pressured contexts is not sophistication — it is a failure mode that produces decision paralysis or delayed action.

---

5. Tacit Knowledge Requires Structured Elicitation

Experts cannot directly report how they decide. But structured retrospective interviews — timeline reconstruction, progressive deepening, hypothetical probing ("what would have changed your read?") — can surface the cues, heuristics, and mental models they cannot otherwise articulate.

Agent implication: You cannot build an expert agent from explicit rules alone. The knowledge that matters is the knowledge experts don't know they have. Eliciting it requires specific, multi-pass methodology.

---

Decision Frameworks

Should I recognize-and-act or deliberate-and-compare?

Is the situation pattern-matchable to prior experience?
  YES → Generate one plausible action. Mentally simulate consequences.
         Simulation passes? → Act.
         Simulation fails? → Modify action or generate next candidate.
  NO  → Is this a genuine novelty or a failure of pattern library?
         Genuine novelty → Deliberate carefully. Document for future pattern addition.
         Pattern library gap → Flag as training/knowledge gap. Proceed with explicit uncertainty.

---

The Quick Test: When to think vs. when to act

Is delay acceptable given time pressure?     NO → Act on best current read
Is the cost of error catastrophic?           YES → Think longer before acting
Is the situation unfamiliar or atypical?     YES → Slow down; situation assessment first
Are cues conflicting or incomplete?          YES → Seek targeted information; don't act blind

See: references/satisficing-and-the-quick-test.md

---

Diagnosing a decision failure

Did the action fail because the situation was misread?
  YES → Failure is in SITUATION ASSESSMENT. Review cue interpretation, not action selection.
Did the action fail because the wrong action was chosen from a correct read?
  YES → Failure is in ACTION SELECTION. Review option generation or simulation quality.
Did the agent act before completing situation assessment?
  YES → Failure is in PREMATURE CLOSURE. Novice or Competent-stage behavior.
Did the agent keep analyzing past the point where action was needed?
  YES → Failure is in DECISION PARALYSIS. Satisficing rules were absent or violated.

See: references/failure-modes-in-high-stakes-dynamic-systems.md

---

Calibrating behavior to expertise level

Agent is Novice or Advanced Beginner?
  → Follow rules strictly. Do not improvise. Flag uncertainty explicitly.
Agent is Competent?
  → Rules apply. Beginning to see exceptions — but do not act on exceptions
    without explicit situational justification. This is the danger zone.
Agent is Proficient or Expert?
  → Pattern recognition is reliable. Intuitive reads are trustworthy signals.
    Anomalies in intuition should be examined, not overridden automatically.

See: references/expert-vs-novice-decision-signatures.md

---

Building a knowledge base from expert knowledge

Start with: Simulate realistic scenarios (don't use retrospective verbal reports alone)
Then: Structured CDM interview — timeline reconstruction → progressive deepening → hypotheticals
Extract: Cues noticed, mental models applied, decision points, what would have changed the read
Encode as: Pattern + Situation Model + Prediction + Action + Simulation result
Validate: Have other experts challenge the encoded model

See: references/tacit-knowledge-elicitation-for-agent-knowledge-bases.md

---

Reference Files

| File | Load When | |------|-----------| | references/recognition-primed-decision-making-for-agents.md | You need to understand or implement the RPD architecture: pattern → action → simulation → revise | | references/expert-vs-novice-decision-signatures.md | You need to assess expertise level, diagnose stage-specific failures, or design staged training | | references/satisficing-and-the-quick-test.md | You need to know when to stop deliberating, or you're designing stop-search rules for an agent | | references/situation-assessment-as-the-critical-skill.md | Failures seem to involve correct actions on wrong situations, or you need to improve diagnosis quality | | references/tacit-knowledge-elicitation-for-agent-knowledge-bases.md | You're building an agent knowledge base from human expertise, interviewing experts, or designing evaluation scenarios | | references/failure-modes-in-high-stakes-dynamic-systems.md | Post-incident analysis, understanding why a decision went wrong, anticipating failure modes by system design | | references/uncertainty-acceptance-and-information-seeking.md | Operating under incomplete information, deciding when more information is worth seeking vs. acting on current read |

---

Anti-Patterns

These are the mistakes this framework specifically warns against:

1. Option enumeration as default Generating all possible actions before choosing is a novice pattern mistaken for rigor. Under time pressure, it is a failure mode. Experts don't enumerate — they recognize and simulate.

2. Treating situation assessment as a formality Jumping to action selection before the situation is adequately understood. The action may be technically correct and still fail because the read was wrong.

3. Competent-stage overconfidence Beginning to deviate from rules without having the situational models that make rule-deviation safe. The most dangerous expertise level is Competent — enough experience to improvise, not enough to know when improvisation is warranted.

4. Unbounded optimization Treating "find the best solution" as always superior to "find a good-enough solution quickly." In dynamic, time-pressured conditions, optimization that delays action is a form of failure.

5. Eliciting expert knowledge through direct self-report Asking experts "how do you decide?" and taking the answer at face value. Experts' verbal accounts of their own cognition are systematically incomplete. The knowledge that matters is tacit and requires structured extraction methodology.

6. Treating uncertainty as a blocker Waiting for complete information before acting. Expert systems operate under irreducible uncertainty and must act on probabilistic situation models while remaining alert to disconfirming cues.

7. Assuming failure is in action selection When reviewing bad outcomes, defaulting to "we chose the wrong action." Most expert-level failures are in situation assessment — the action was reasonable given the (incorrect) read.

---

Shibboleths

How to tell if someone has actually internalized this framework — versus just having read about it.

They say "what was the situation assessment?" before "what action did they take?" Someone who has internalized NDM goes to diagnosis before judgment. Novice thinkers evaluate the action; expert thinkers first ask whether the situation was correctly read.

They distinguish between "not enough options generated" and "wrong situation model" Surface readers conflate decision failure types. Internalized thinkers immediately ask: was this a failure of recognition, simulation, or situation assessment?

They treat satisficing as principled, not as settling Naive readers see satisficing as "doing your best with limited time." Internalized thinkers understand it as a rational strategy with explicit stop-search logic — not a concession to imperfection.

They are suspicious of expert self-reports Someone who has really absorbed this framework knows that what experts say they do and what they actually do are systematically different. They design elicitation processes accordingly — they don't just interview.

They can name what's dangerous about the Competent stage Novice → Expert is not a monotonic improvement in safety. The Competent stage is specifically hazardous because rule-breaking begins before situational models are rich enough to justify it. This is non-obvious and only retained through genuine engagement.

They ask "what cues was the agent responding to?" before evaluating the response The cue-interpretation-action chain is central to NDM. Someone who has internalized the framework diagnoses from cues forward, not from outcomes backward.

They treat intuition as data, not noise Expert intuition (feeling that something is wrong, that a situation is atypical) is the output of unconscious pattern-matching against rich experience. Internalized NDM practitioners treat strong intuitions as signals requiring investigation — not as irrationality to be suppressed.

---

Load reference files on demand. This index is the activation point — go deeper only where the problem demands it.