curiositech/sensemaking-focus-model

SKILL: FOCUS Sensemaking Model

Core Concept

Build agent reasoning systems that mirror how experts develop situation understanding through continuous frame-data reciprocity, fragmentary mental models, and assumption-challenging questions rather than comprehensive world models or linear information processing.

Decision Points

When Agent Encounters Anomaly

IF anomaly_count < 2 AND current_frame_confidence > 0.7
  → USE Preserving (explain within current frame)
  → SET anomaly_accumulation_threshold = 3
  → CONTINUE with current frame

ELIF anomaly_count >= 3 OR core_expectation_violated
  → FORCE Comparing function
  → GENERATE alternative frames from FMM library
  → REQUIRE explicit frame comparison before proceeding
  
ELIF anomaly is assumption-critical
  → IMMEDIATELY trigger Questioning
  → EXTRACT assumptions underlying current frame
  → GENERATE assumption-challenging questions
  → IF no satisfactory answer → move to Re-framing

When Building Agent Knowledge Base

IF task requires expert-level performance
  → BUILD fragmentary mental models (FMMs):
    - Local cause-effect patterns
    - Cue-pattern associations  
    - Context-dependent rules
    - Domain-specific heuristics
  → AVOID comprehensive world models

ELIF system needs to improve over time
  → SEQUENCE: Knowledge structures first, then metacognition
  → NOT generic "critical thinking" strategies

When Agent Must Generate Questions

IF current frame has high confidence
  → EXTRACT core assumptions from frame
  → PRIORITIZE assumption-challenging over information-seeking
  → ASK: "Why would X do Y?" not "What did X do?"

ELIF multiple plausible frames exist
  → GENERATE discriminating questions
  → FOCUS on questions that would invalidate one frame
  → TARGET fundamental relationships, agency, causality

Failure Modes

1. Fixation Through Over-Preservation

Detection Rule: If agent explains away 4+ anomalies without triggering frame comparison Symptoms:

• Anomaly count climbing while frame confidence stays high
• Increasingly complex explanations for simple contradictions
• Resistance to considering alternative interpretations Fix: Implement hard threshold - after N anomalies, FORCE Comparing function regardless of confidence

2. Comprehensive Model Brittleness

Detection Rule: If system breaks on edge cases not in training data OR requires complete domain knowledge before functioning Symptoms:

• Analysis paralysis while trying to build complete understanding
• System failure when encountering novel scenarios
• Inability to function with partial information Fix: Decompose into FMM library with just-in-time assembly patterns

3. Sensemaking-Decision Blur

Detection Rule: If agent proposes actions while situation assessment confidence < 0.6 Symptoms:

• Action recommendations changing as new data arrives
• Inability to explain "what situation we're in" separately from "what we should do"
• Premature commitment to courses of action Fix: Architectural separation with explicit confidence gate between phases

4. Information-Only Questioning

Detection Rule: If 80%+ of generated questions are "What/Where/When" rather than "Why/Who decides/What if" Symptoms:

• Questions like "What is X?" instead of "Why would X happen?"
• Data-seeking without assumption-challenging
• Missing critical insights about agency and causality Fix: Implement assumption-extraction module before question generation

5. Linear Processing Pipeline

Detection Rule: If system processes data → analysis → conclusion without feedback loops Symptoms:

• Current frame doesn't guide what information to seek next
• New data doesn't trigger frame revision
• No reciprocal relationship between interpretation and observation Fix: Rebuild as continuous loop where frame guides data selection and data challenges frame

Worked Examples

Example 1: Intelligence Analysis Anomaly

Scenario: Agent analyzing communication patterns notices 40% drop in message volume from Target A Initial Frame: "Target A reducing operational tempo" Confidence: 0.8

Decision Point Navigation:

1. Anomaly Detection: Drop contradicts recent intelligence about planned operations
2. Threshold Check: First major anomaly, frame confidence high → Try Preserving
3. Preserving Attempt: "Maybe delayed operations due to weather"
4. New Anomaly: Weather data shows clear conditions
5. Anomaly Count: Now at 2, still below threshold
6. Third Anomaly: Increase in communication from Target B (previously dormant)
7. Trigger: anomaly_count = 3 → FORCE Comparing function

Frame Comparison:

• Frame A: "Reduced tempo" (explains volume drop, conflicts with weather/B activity)
• Frame B: "Communication security change" (explains A drop + B increase)
• Frame C: "Target A compromised/offline" (explains all data)

Expert vs Novice: Novice would continue explaining anomalies within Frame A. Expert triggers comparison at threshold, generates assumption-challenging questions: "Why would B activate when A reduces?" "Who decides communication protocols?"

Example 2: System Performance Diagnosis

Scenario: Application response time increased 300% overnight Initial Frame: "Database performance degradation"

FMM Assembly Process:

• Pattern: "Sudden overnight change" → Check deployment logs first
• Pattern: "3x performance drop" → Usually resource contention or config change
• Pattern: "Database suspected" → Verify with query time metrics

Questioning Strategy:

• Information-seeking: "What changed in database?"
• Assumption-challenging: "Why assume it's database? What if network latency increased?"
• Agency-focused: "Who has access to modify configs overnight?"

Frame Revision: Questions reveal recent load balancer configuration change, not database issue. FMM pattern "config change + performance drop" provides new frame.

Example 3: Market Analysis Under Uncertainty

Scenario: Stock showing unusual trading patterns - high volume, price stability Multiple Competing Frames:

• Frame A: "Institutional accumulation"
• Frame B: "Insider trading before announcement"
• Frame C: "Algorithm testing by trading firm"

Assumption-Challenging Questions:

• Frame A assumption: "Institutions want to accumulate quietly" → "Why wouldn't they use dark pools?"
• Frame B assumption: "Insiders have material information" → "Why risk detection with high volume?"
• Frame C assumption: "Testing requires real market" → "Why not use sandbox environments?"

Trade-off Documentation:

• Preserving Frame A: Low cognitive cost, but may miss real insider activity
• Comparing all frames: High cognitive cost, but better anomaly coverage
• Decision: Set 48-hour monitoring period with daily frame reassessment

Quality Gates

Sensemaking Phase Completion

• [ ] Current frame explains 80%+ of observed data without forced explanations
• [ ] Anomaly accumulation count < threshold (typically 3)
• [ ] Frame confidence score > 0.6 on established scale
• [ ] At least 2 alternative frames considered and explicitly rejected
• [ ] Core assumptions of current frame identified and documented

Knowledge Base Adequacy

• [ ] FMM library contains 50+ local cause-effect patterns for domain
• [ ] Cue-pattern associations validated against expert examples
• [ ] Context-dependent rules specify activation conditions
• [ ] Just-in-time assembly procedures tested on novel scenarios
• [ ] Pattern coverage verified against known expert reasoning traces

Question Generation Quality

• [ ] 60%+ of questions challenge frame assumptions vs. seek information
• [ ] Questions target agency ("Who decides?"), causality ("Why would X?"), relationships
• [ ] Generated questions would discriminate between competing frames
• [ ] Assumption-extraction module identifies 3+ testable frame dependencies
• [ ] Questions match expert questioning patterns for domain

System Architecture Validation

• [ ] Explicit separation between sensemaking and decision-making phases
• [ ] Confidence threshold gate prevents premature action planning
• [ ] Frame-data reciprocity loop implemented with bidirectional updates
• [ ] Anomaly accumulation triggers automatic frame comparison
• [ ] System can operate effectively with incomplete information

Anti-Pattern Prevention

• [ ] Hard limits prevent excessive anomaly explanation (Preserving fixation)
• [ ] Architectural constraints prevent sensemaking-decision blur
• [ ] FMM approach prevents comprehensive model brittleness
• [ ] Question generation avoids information-only patterns
• [ ] Continuous loops prevent linear processing pipeline

Not-For Boundaries

Do NOT use this skill for:

• Simple classification tasks → Use pattern-matching algorithms instead
• Complete information scenarios → Use standard decision trees instead
• Real-time response systems → Use fast heuristics, delegate to reaction-systems
• Well-structured problems → Use domain-specific optimization instead
• Single-frame situations → Use direct reasoning, no frame comparison needed

Delegate to other skills:

• For mathematical optimization → Use operations-research methods
• For known problem categories → Use domain-specific algorithms
• For collaborative sensemaking → Use group decision-making frameworks
• For time-critical decisions → Use rapid decision-making protocols
• For routine monitoring → Use anomaly detection without frame revision

Clear boundary markers:

• If problem has single correct answer discoverable through analysis → Not sensemaking
• If situation is familiar with established procedures → Not sensemaking
• If speed matters more than understanding → Not sensemaking
• If you need group consensus rather than individual understanding → Use collaborative methods