Swiss Cheese Validation: Verify Critic Independence

Purpose

Verify that multiple critic layers have orthogonal failure modes, ensuring the Swiss cheese model is working correctly. Measure overlap in critic rejections and warn if critics catch the same errors.

Theory

The Swiss cheese model achieves reliability through misaligned holes:

•Each critic has failure modes (errors it misses)
•Holes don't align across critics
•Errors caught by different critics don't overlap significantly

Target: Independence score > 0.85 (less than 15% overlap)

Validation Process

Measure Overlap

Analyze rejection patterns across sessions:

python

code_rejections = sessions_where_code_critic_vetoed
security_rejections = sessions_where_security_critic_vetoed

overlap = code_rejections ∩ security_rejections
union = code_rejections ∪ security_rejections

independence_score = 1 - (|overlap| / |union|)

Interpretation:

•Score > 0.90: Excellent independence (orthogonal)
•Score 0.80-0.90: Good independence (acceptable)
•Score < 0.80: Poor independence (critics overlapping)

Expected Results

From research paper (522 sessions):

Critic Pair	Overlap	Independence
Code vs Security	2.3%	0.977 ✓
Code vs Domain	3.1%	0.969 ✓
Security vs Domain	0.8%	0.992 ✓

Warning Indicators

High overlap detected when:

•Same sessions rejected by multiple critics
•Similar rejection reasons across critics
•Independence score < 0.80

Causes:

•Critics using same model/provider (cognitive monoculture)
•Similar training data
•Overlapping validation logic

Solutions:

•Use different model providers per critic
•Adjust critic specializations
•Review critic prompts for duplication

Configuration

Layer Definition

Default (hardcoded):

yaml

layers:
  - code-critic
  - security-critic
  - domain-critic

Custom (user-configured):

yaml

layers:
  - code-critic
  - security-critic
  - financial-critic
  - performance-critic

Dynamic (context-based):

yaml

layers:
  financial_files: [code-critic, financial-critic]
  auth_files: [code-critic, security-critic]
  api_files: [code-critic, performance-critic]

Validation Report

code

SWISS CHEESE VALIDATION REPORT

Analysis Period: Last 30 days (87 sessions)

INDEPENDENCE SCORES:
✓ Code vs Security: 0.981 (excellent)
✓ Code vs Domain: 0.973 (excellent)
✓ Security vs Domain: 0.995 (excellent)

OVERLAP ANALYSIS:
- Code & Security both rejected: 2 sessions (2.3%)
  Reasons: Authentication logic had both syntax error AND security flaw

- Code & Domain both rejected: 3 sessions (3.4%)
  Reasons: Financial calculation had logic error AND violated precision rules

- Security & Domain both rejected: 1 session (1.1%)
  Reasons: Data validation missing AND violated GDPR requirements

CONCLUSION: ✓ Swiss cheese model working correctly
All critics demonstrate orthogonal failure modes.

When Validation Fails

code

⚠ INDEPENDENCE CONCERN DETECTED

Analysis: Last 30 days (87 sessions)

POOR INDEPENDENCE:
❌ Code vs Security: 0.72 (poor - 28% overlap)

ISSUE: Both critics rejecting same sessions frequently
Sample overlaps:
- Session #42: Both caught missing input validation
- Session #55: Both caught SQL injection in user input
- Session #61: Both caught unescaped HTML in output

ROOT CAUSE: Security critic checking code-level details
RECOMMENDATION: Refocus security critic on OWASP patterns only,
let code critic handle general logic validation

Active Monitoring

Run validation automatically after every 20 sessions:

yaml

monitoring:
  frequency: every_20_sessions
  alert_threshold: 0.80
  action: warn_user

Additional Resources

•references/independence-verification.md - Statistical methods for measuring orthogonality
•references/failure-mode-analysis.md - Common causes of critic overlap
•examples/poor-independence.md - Case study of overlapping critics and remediation