curiositech/ndm-decision-models

Naturalistic Decision Making Models for Agent Systems

Decision Points

Primary Branch: Time Pressure + Domain Familiarity → Decision Mode

Agent faces decision task
├── High time pressure (< 5 seconds to decide)
│   ├── Domain is familiar/trained → Use RPD mode
│   │   ├── Clear situation pattern? → Generate action, simulate, execute
│   │   └── Unclear pattern? → Generate best guess action, short simulation, act
│   └── Domain is novel/untrained → Use constrained analytical mode
│       ├── Can identify 2-3 viable options quickly? → Compare those only
│       └── Cannot quickly identify options? → Escalate to human/expert
└── Low time pressure (> 30 seconds to decide)
    ├── High stakes + reversible decision → Use RPD with extended simulation
    ├── High stakes + irreversible decision → Use analytical mode with expert review
    └── Low stakes → Use RPD mode regardless of domain familiarity

Secondary Branch: Action Validation During Execution

Agent generates action via RPD
├── Mental simulation passes cleanly → Execute immediately
├── Mental simulation shows minor issues → Modify action, simulate again
├── Mental simulation shows major failure → Generate different action
└── Cannot simulate (insufficient domain model) → Revert to analytical mode

Tertiary Branch: Multi-Agent Coordination

Multiple agents must coordinate
├── Agents have shared situation model → Proceed with individual RPD
├── Agents disagree on situation assessment → Stop, build shared model first
└── Situation model unclear → Designate lead agent for situation assessment

Failure Modes

1. Analysis Paralysis in Pattern-Recognizable Situations

Detection Rule: Agent spends >10 seconds comparing obvious alternatives when first option would work. Symptoms: Over-enumeration of options, probability calculations for clear cases, delayed response to time-critical situations. Fix: Check if situation matches trained patterns. If yes, force RPD mode; generate first workable action and execute after brief simulation.

2. Recognition Bias in Novel Domains

Detection Rule: Agent confidently executes actions in unfamiliar domains without simulation or verification. Symptoms: Fast decisions in areas outside training data, no uncertainty signaling when domain shifts, pattern matching to superficially similar but structurally different situations. Fix: Add domain boundary detection. When domain novelty detected, require analytical mode or human consultation.

3. Simulation Bypass Under Pressure

Detection Rule: Agent executes first generated action without mental simulation when time pressure increases. Symptoms: Higher error rates under time pressure, no modification of initially generated actions, inability to catch obvious flaws in plan. Fix: Implement minimum simulation requirement even under extreme time pressure. Better to act 2 seconds later with simulation than immediately without.

4. Situation Model Lock-in

Detection Rule: Agent maintains initial situation assessment despite contradictory evidence emerging during execution. Symptoms: Continued execution of failing plan, ignoring feedback that invalidates situation model, escalating commitment to wrong diagnosis. Fix: Build expectancy violation monitoring. Force situation reassessment when 2+ predictions fail to materialize.

5. Decision Support Tool Override

Detection Rule: Agent ignores or works around decision support tools that require analytical processing. Symptoms: Consistent bypass of formal decision frameworks, resistance to using probability estimation tools, degraded performance when tools are mandatory. Fix: Redesign tools to support situation assessment and pattern recognition rather than option comparison.

Worked Examples

Example 1: Emergency Response Agent - RPD Success

Scenario: Building fire alarm triggers emergency response agent. Sensors show: smoke detector C-wing, temperature spike, no water flow alerts, 14:30 weekday.

Agent's RPD Process:

1. Recognition: Pattern matches to "office fire, business hours, sprinkler system intact"
2. Action Generation: Evacuate C-wing, dispatch fire crew, prepare building-wide evacuation
3. Mental Simulation: C-wing evacuation takes ~3 minutes, fire crew arrival ~7 minutes, if fire spreads beyond C-wing need full evacuation. Simulation passes.
4. Execution: Issues C-wing evacuation, dispatches crew

What Novice Would Miss: Would spend time calculating probability fire spreads, comparing evacuation vs. wait-and-see options, analyzing sensor readings. By recognition, expert agent knows: office fire + working sprinklers = evacuate zone first, full building only if escalation.

Outcome: Fire contained to origin room. Total evacuation time: 4 minutes. Analytical approach would have taken 8-12 minutes just for decision.

Example 2: Trading Agent - Recognition Failure

Scenario: Market volatility spike during Asian session. Agent sees pattern similar to "flash crash" from training data: rapid 2% drop in 5 minutes, high volume, news feed shows "regulatory concern."

Agent's Flawed RPD:

1. Recognition: Matches to "flash crash recovery" pattern
2. Action Generation: Buy the dip, expecting rapid rebound
3. Mental Simulation: Predicts 1-2% recovery within 30 minutes
4. Execution: Takes large long position

What Expert Would Catch: "Regulatory concern" during Asian session is structurally different from technical flash crashes. Domain shift not recognized. Should have triggered analytical mode or expert consultation.

Outcome: Further 3% drop as regulatory news proves substantial. Loss: $2.3M.

Fix Applied: Added domain boundary detection for "regulatory news" keyword that forces analytical mode regardless of price pattern recognition.

Example 3: Multi-Agent Coordination - Situation Model Disagreement

Scenario: Software deployment agents preparing production release. Agent A sees "standard deployment" pattern, Agent B sees "high-risk deployment" pattern from same signals: 47 code changes, 3 database migrations, 2 new external dependencies, Friday 4PM release window.

Decision Process:

1. Detect Disagreement: Agents generate different actions (A: proceed normally, B: delay to Monday)
2. Stop Individual RPD: Both agents halt action generation
3. Build Shared Model: Agent A weights "only 47 changes, tested migrations"; Agent B weights "Friday release + external deps"
4. Resolution: Shared assessment: "Standard scope but risky timing"
5. Coordinated Action: Proceed with deployment but extend monitoring window and prepare rapid rollback

What Would Fail: If agents proceeded with individual RPD, would get coordination failure. Agent A deploys while Agent B holds back monitoring resources.

Quality Gates

Agent deployment readiness checklist:

• [ ] Agent correctly identifies domain boundaries and switches decision modes appropriately
• [ ] Agent can generate workable actions within time constraints for 90%+ of trained scenarios
• [ ] Agent performs mental simulation before execution and catches major failure modes
• [ ] Agent updates situation assessment when initial predictions fail (expectancy violation response)
• [ ] Agent escalates to analytical mode or human consultation when facing novel/untrained situations
• [ ] Agent coordinates effectively with other agents by establishing shared situation models first
• [ ] Agent's first-generated action is workable (not necessarily optimal) in 85%+ of test cases
• [ ] Agent's decision latency is appropriate for time pressure context (fast for urgent, thorough for high-stakes)
• [ ] Agent ignores or bypasses decision support tools that interfere with recognition-based processing
• [ ] Agent maintains performance under time pressure without complete simulation bypass

NOT-FOR Boundaries

This skill should NOT be used for:

• Novel problem domains: Use analytical-reasoning instead for completely unfamiliar situations
• Single high-stakes irreversible decisions: Use formal-decision-analysis for "bet the company" choices
• Mathematical optimization problems: Use optimization-algorithms for resource allocation, scheduling, routing
• Regulatory compliance decisions: Use rule-based-systems for legal/compliance requirements where process traceability is mandatory
• Creative/generative tasks: Use design-thinking or creative-problem-solving for open-ended innovation
• Research and discovery: Use scientific-method for hypothesis testing and knowledge creation

Delegate instead:

• For mathematical problems with clear optimization criteria → optimization-algorithms
• For creative ideation → design-thinking
• For regulatory compliance → rule-based-systems
• For novel domains with no training data → analytical-reasoning
• For high-stakes irreversible decisions → formal-decision-analysis