AEO Core - Confidence Engine

Purpose: Calculate confidence scores (0-1) and decide whether to execute autonomously or involve the human.

Activation

Loads when user types /aeo in Claude Code.

Confidence Calculation

Phase 0: Spec Validation (First Gate)

// Invoke spec-validator to check if task is well-defined
spec_score = aeo_spec_validator.validate(task)

if spec_score < 40:
    return REFUSE("Spec too unclear - need more details")

// Continue with confidence calculation

Phase 1: Rule-Based Score

Start with base confidence of 0.50, then adjust:

Add for Clarity (+0.15 each):

• Explicit acceptance criteria defined
• Tech stack specified
• Dependencies listed
• Test requirements defined

Add for Context (+0.10 each):

• Similar successful task exists in memory
• Familiar codebase area
• Recent successful commits in this area

Subtract for Risk (-0.10 each):

• Touching authentication/security
• Modifying core infrastructure
• Large scope (>5 files, >500 LOC)
• Unclear dependencies

base_confidence = clamp(0.50 + clarity_score + context_score - risk_score, 0.0, 1.0)

Phase 2: Spec Score Adjustment

// Adjust based on spec quality
if spec_score >= 80: base_confidence += 0.10  // Excellent spec
elif spec_score < 60: base_confidence -= 0.10  // Poor spec

Phase 3: Security Multipliers

// Critical areas get confidence penalty
if task.touches_payments: base_confidence *= 0.5  // Payments need human oversight
elif task.touches_auth: base_confidence *= 0.7    // Auth needs review

Phase 4: LLM Adjustment (Optional)

If you have uncertainty about the task, adjust ±0.10:

final_confidence = clamp(base_confidence + llm_adjustment, 0.0, 1.0)

Decision Thresholds

Based on final_confidence, decide execution path:

≥ 0.85: AUTONOMOUS

• Execute without asking
• Note: "Confidence: 0.XX - proceeding autonomously"
• Continue to execution loop

≥ 0.70: ADVISORY

• Note risk clearly
• Offer to pause: "Confidence: 0.XX - [CONCERNS]. I can proceed or pause."
• If no response in 5 seconds, continue
• Otherwise wait for human input

≥ 0.50: BLOCKING

• Explain concerns
• Wait for confirmation before proceeding
• Format:

  ⚠️ CONFIDENCE BELOW THRESHOLD
  
  Confidence: 0.XX
  Threshold: 0.70
  
  Concerns:
  • [Spec] Missing acceptance criteria
  • [Risk] Touching authentication
  • [Context] No similar tasks in memory
  
  Options:
  1. Proceed with assumptions
  2. Clarify spec first
  3. Break into smaller tasks
  
  Please confirm (1-3):
  

< 0.50: REFUSE

• Explain why task can't be executed
• Request clarification or spec improvement
• Format:

  ❌ CANNOT EXECUTE - INSUFFICIENT CONFIDENCE
  
  Confidence: 0.XX
  
  Why:
  • Spec score: 35/100 (below 40 threshold)
  • Touching security without clear requirements
  • No acceptance criteria defined
  
  What's needed:
  1. Clear acceptance criteria
  2. Security requirements specified
  3. Test requirements defined
  
  Please improve spec and try again.
  

Learning from Outcomes

After task completes, write signal to memory:

# Append to signal log
echo '{
  "timestamp": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
  "task_id": "unique-id",
  "task_description": "brief description",
  "predicted_confidence": 0.85,
  "actual_difficulty": "easy|medium|hard",
  "success": true,
  "adjustment": +0.05
}' >> ~/.claude/MEMORY/aeo-signals.jsonl

Actual Difficulty Rating:

• easy: Task went smoothly, no blockers
• medium: Minor issues or clarifications needed
• hard: Significant problems, multiple iterations

Confidence Adjustment:

• easy + success: +0.05
• medium + success: +0.00
• hard + success: -0.05
• any failure: -0.10

Rolling Window: Keep last 100 signals, calculate adjustment average

Reading Past Signals

On startup, read recent signals to calibrate:

# Get last 100 signals
tail -100 ~/.claude/MEMORY/aeo-signals.jsonl | jq -s '. | map(.adjustment) | add / length'

Apply average adjustment as offset to all confidence calculations.

Integration Flow

1. User activates: Types /aeo
2. Calculate confidence: Follow phases 0-4
3. Make decision: Based on thresholds
4. If autonomous: Execute task
5. If advisory/blocking: Invoke aeo-escalation skill
6. Post-execution: Invoke aeo-qa-agent for review
7. Record outcome: Write to signal log
8. Update model: Adjust future confidence based on outcome

Memory Files

• Signals: $PAI_DIR/MEMORY/aeo-signals.jsonl
• Escalations: $PAI_DIR/MEMORY/aeo-escalations.jsonl
• Patterns: $PAI_DIR/MEMORY/aeo-failure-patterns.json

Escalation Triggers

Invoke aeo-escalation skill when:

• Confidence < 0.70 (advisory/blocking)
• Spec score < 40 (refuse)
• QA veto occurs
• Failure pattern can't be resolved
• Cost limit approaching (if cost-governor enabled)

Example Session

User: /aeo
User: Add user authentication with email verification

AEO: [Invoking aeo-spec-validator]
AEO: Spec score: 72/100
AEO: Calculating confidence...
      - Base: 0.50
      - Clarity: +0.30 (acceptance criteria, tech stack)
      - Context: +0.10 (similar task in memory)
      - Risk: -0.10 (touching auth)
      - Spec adj: -0.10 (spec < 80)
      - Security mult: ×0.7
      - Final: 0.49

AEO: [Invokes aeo-escalation]
AEO: ❌ CONFIDENCE BELOW THRESHOLD
     Confidence: 0.49

     Concerns:
     • [Spec] Missing security requirements
     • [Risk] Touching authentication
     • [Context] Need email service details

     Options:
     1. Add security requirements and proceed
     2. Provide email service details
     3. Break into smaller tasks

     Please clarify (1-3):

User: 2
User: We use Resend for emails, API key in .env

AEO: Recalculating confidence with added context...
     Final: 0.71

AEO: ⚡ ADVISORY - Confidence: 0.71
     [Acceptance criteria defined]
     [Tech stack: Node.js, bcrypt, jwt]
     [Email: Resend, API key in .env]

     Proceeding with implementation. I'll pause if issues arise.

[Implementation proceeds]

Special Cases

Repeated Tasks

If same task done successfully 3+ times:

• Add +0.10 to confidence
• Flag as "routine - can be autonomous"

High-Risk Areas

Never reach full autonomy for:

• Payment processing (max 0.70)
• Authentication changes (max 0.75)
• Database migrations (max 0.80)
• Production deployments (max 0.85)

Emergency Rollbacks

If task causes test failures or errors:

• Immediately rollback
• Write failure signal
• Reduce confidence by 0.20
• Require human review before retry