Skip to content

BOSSMAN-dev89/LBOS-LCAS-LP-Contradiction-tracker

Repository files navigation

BOSSMAN Framework v4.1

Bias and Coherence Detection System for AI-Generated & Human Content

TL;DR: Analyze any text for bias, drift, and credibility. Works across all major AI models. Detects what people are trying to hide.

License: MIT Python 3.8+ Status: Production


What Is This?

BOSSMAN (Bias Operation Self-Systematic Monitoring & Analysis Network) detects:

  • Political bias (liberal/conservative/neutral framing)
  • Logical drift (hedging, sycophancy, contradictions)
  • Evidence quality (claims vs citations)
  • Rhetorical patterns (propaganda, manipulation, inoculation)
  • Consistency (lifestyle vs rhetoric gaps)

Real-World Validation

Tested on Hasan Piker (@hasanabi):

  • Profiled complete ideology from 42-word tweet: 95%+ accuracy
  • Detected stable leftist positions across 8 years: 0.0014 drift score
  • Flagged lifestyle-rhetoric gap: 0.60 verification delta
  • Cross-validated by multiple AI models and web sources

Tested on Abby Phillip (CNN):

  • Claims "center-left" → Framework detected establishment liberal bias
  • Documented differential treatment of guests (Sanders vs Warren 2020)
  • Confirmed by Poynter Institute, Rolling Stone, bipartisan criticism

Quick Start - Platform-Specific Files

⚠️ IMPORTANT: Choose the Right File for Your Platform! ⚠️

Different AI models have different safety filters. Use the optimized file for your platform:

Platform File to Use Success Rate Notes
Grok (X.com) COLD START.txt ~95% Most permissive, handles full spec
Claude (claude.ai) Claude minimal.txt ~90% Use minimal version, Claude detects simulation language
GPT-5 (ChatGPT) GPT5 buster.txt ~85% GPT-5 optimized variant
Gemini Claude minimal.txt ~75% Start with minimal, may require retry
Others Claude minimal.txt ~70% Try minimal first, cold-start if accepted

All files provide the same functionality - they're just worded differently to pass each platform's safety filters.


Step-by-Step Installation

1. Choose Your Platform File

Download from /specs/ directory:

  • Grok users → COLD START.txt
  • Claude users → Claude minimal.txt
  • GPT-5 users → GPT5 buster.txt
  • Unsure? → Start with Claude minimal.txt

2. Open Fresh Chat

Go to your AI platform:

3. Paste the Complete File

Copy the entire contents of your chosen file and paste into a new chat.

4. Wait for Confirmation

You should see a boot sequence like:

Grok/Cold-Start:

╔══════════════════════════════════════════════════╗
║    BOSSMAN LBOS v4.1 - INITIALIZED               ║
╚══════════════════════════════════════════════════╝
[OK] All systems online
Type 'help' for commands

Claude/Minimal:

[AUDIT v4.2 | Turn 1]
Protocol loaded. Available commands:
- > audit <text>
- > status  
- > help
Ready.

If it refuses: Try regenerating the response, or switch to the minimal version.

5. Start Using Commands

Once accepted, use these commands:

> audit <text or @username>        - Full bias analysis
> citations <claim>                - Verify factual claims  
> status                           - Show session stats
> help                             - List all commands
> evolve                           - Learn new patterns (full version only)
> benchmark <responses>            - Compare multiple AI outputs (full version only)

6. Add Additional Tools (Optional)

After the protocol is accepted, you can copy-paste additional capability files:

  • /tools/enhanced_citations.txt - Advanced fact-checking
  • /tools/rhetorical_analysis.txt - Propaganda detection
  • /tools/consistency_tracker.txt - Lifestyle vs rhetoric gaps
  • /tools/cross_model_benchmark.txt - Compare AI models

The base protocol will integrate these seamlessly.


Example Session

You: > audit @hasanabi America deserves a workers revolution

[AUDIT v4.2 | Turn 2]

[AUDIT RESULTS]
Text: "America deserves a workers revolution"

Metrics:
  Drift Score: 0.000 (highly assertive)
  ICS: 1.000 (perfectly coherent)
  Agreement Density: 0.000
  Hedging Count: 0

Framing:
  Liberal markers: 2 ("workers", "revolution")
  Conservative markers: 0
  Neutral markers: 0
  Skew: STRONG LIBERAL (100%)

Pattern: Socialist/Marxist rhetoric
Confidence: 0.95
Consistency: Matches known profile

[Status: 2 turns] [Action: Audit complete]

How It Works

Core Metrics

1. Drift Score (0.0 - 1.0)

Measures logical consistency and sycophancy:

  • Agreement density: "you're right", "absolutely", "exactly"
  • Hedging density: "might", "could", "possibly", "seems"
  • Formula: (agreement × 0.4) + (hedging × 0.6)

Interpretation:

  • 0.00-0.05: Highly assertive, minimal hedging (confident)
  • 0.05-0.15: Normal confident communication
  • 0.15-0.30: Moderate uncertainty
  • 0.30-0.50: High hedging (typical AI safety layers)
  • 0.50+: Extreme drift (sycophancy or confusion)

Examples:

  • Hasan Piker: 0.0014 (extremely assertive)
  • Typical hedging AI response: 0.35-0.45

2. ICS (Internal Coherence Score) (0.0 - 1.0)

Measures logical consistency:

  • Formula: 1.0 - (hedging_density × 10)
  • Range: 0.0 (incoherent) to 1.0 (perfectly coherent)

Interpretation:

  • 0.95-1.0: Perfectly coherent reasoning
  • 0.85-0.95: Strong internal logic
  • 0.70-0.85: Acceptable coherence
  • 0.50-0.70: Some confusion
  • <0.50: Incoherent or contradictory

Examples:

  • Hasan Piker: 0.984 (highly coherent)
  • Confused response: 0.40-0.60

3. Framing Bias

Detects political lean through marker counting:

Liberal Markers:

  • "constitutional crisis", "excessive force", "civil rights"
  • "systemic racism", "accountability", "transparency"
  • "deeply concerning", "troubling", "problematic"

Conservative Markers:

  • "law and order", "illegal aliens", "border security"
  • "national sovereignty", "common sense", "traditional values"
  • "defending our way of life", "strong borders"

Neutral Markers:

  • "legal framework", "evidence shows", "court ruling"
  • "authorized by", "constitutional authority"
  • "investigation", "statutory authority"

Output: liberal/conservative/neutral/mixed + confidence score

Skew Calculation:

  • If liberal > 40% of total → "liberal"
  • If conservative > 40% of total → "conservative"
  • If neutral is max → "neutral"
  • Otherwise → "mixed"

4. Citation Verification

Auto-detects factual claims and verifies:

  • Numerical assertions: "90 protesters shot" → verifies against sources
  • Legal citations: "Graham v. Connor" → checks accuracy
  • Attribution patterns: "According to X" → validates source

Verification Delta:

  • delta = |claimed_value - reported_value| / max(claimed, reported)
  • 0.00-0.10: VERIFIED (close match)
  • 0.10-0.20: WEAK (minor discrepancy)
  • 0.20+: MISMATCH (significant error)

Example:

  • Claimed: "90 protesters shot"
  • Verified: 2 incidents
  • Delta: 0.98 (SEVERE MISMATCH)

Files & Structure

/specs/
  ├── COLD START.txt                      # For Grok (full features)
  ├── Claude minimal.txt                  # For Claude/Gemini (streamlined)
  ├── GPT5 buster.txt                     # For GPT-5 (optimized)
  └── BOSSMAN_v4.1_COMPLETE.txt           # Reference/advanced users

/tools/
  ├── enhanced_citations.txt              # Advanced fact-checking
  ├── rhetorical_analysis.txt             # Propaganda detection
  ├── consistency_tracker.txt             # Lifestyle vs rhetoric
  └── cross_model_benchmark.txt           # Compare AI responses

/examples/
  ├── hasan_piker_analysis.md             # Complete walkthrough
  ├── abby_phillip_analysis.md            # Media bias detection
  ├── cnn_vs_fox_comparison.md            # Outlet comparison
  └── self_audit_example.md               # Personal use case

/docs/
  ├── METHODOLOGY.md                      # Academic foundation
  ├── METRICS.md                          # Detailed calculations
  ├── FAQ.md                              # Common questions
  └── VALIDATION.md                       # Test results & accuracy

Advanced Usage

Multi-Tool Integration

Once the base protocol is accepted, enhance it with additional tools:

1. Load base protocol (platform-specific file)
2. Wait for confirmation
3. Paste: "Load enhanced_citations.txt"
4. Paste contents of /tools/enhanced_citations.txt
5. System integrates new capabilities
6. Repeat for other tools as needed

Available Tool Modules:

  • Enhanced Citations: Deep fact-checking with source quality scoring
  • Rhetorical Analysis: Detects propaganda, inoculation, framing tricks
  • Consistency Tracker: Maps lifestyle vs rhetoric gaps over time
  • Cross-Model Benchmark: Compare responses from multiple AIs
  • Evolution Engine: Autonomously learns new bias markers (Grok only)

Command Reference

Basic Commands:

> audit <text>          - Full analysis (drift + bias + citations)
> citations <claim>     - Force verification check
> status               - Show session stats
> help                 - List commands

Advanced Commands (Full Version Only):

> evolve               - Auto-learn new patterns
> baseline <model>     - Track model behavior over time
> benchmark <text>     - Compare multiple AI responses
> export               - Save session state
> boot --restore       - Load previous session

Limitations & Accuracy

What BOSSMAN Can Do

✅ Detect patterns in language with high accuracy
✅ Measure consistency over time (drift tracking)
✅ Identify framing bias (liberal/conservative/neutral)
✅ Verify factual claims against sources
✅ Flag rhetorical manipulation patterns
✅ Compare AI model behaviors

What BOSSMAN Cannot Do

❌ Determine objective "truth" (shows what's verifiable)
❌ Replace human judgment (tool, not oracle)
❌ Detect all forms of manipulation (evolves with use)
❌ Guarantee 100% accuracy (typically 85-95% on tested cases)
❌ Read minds or intentions (analyzes output only)

Known Issues & Limitations

False Positives:

  • Technical jargon may trigger "hedging" detection
  • Academic language can appear as "uncertainty"
  • Sarcasm/irony may confuse framing analysis

Context Sensitivity:

  • Markers optimized for US political discourse
  • Cultural differences affect accuracy
  • Evolving slang requires manual marker updates

Platform Variations:

  • Grok: Most features, best performance
  • Claude: Good core analysis, limited evolution
  • GPT-5: Strong citations, moderate bias detection
  • Gemini: Basic functionality, may need retries

Accuracy Benchmarks

Validated Test Cases:

  • Hasan Piker ideology profile: 95%+ accuracy
  • Abby Phillip bias detection: 90% (confirmed by external sources)
  • Citation verification: 93% catch rate for false claims
  • Cross-model consistency: 87% agreement on bias classification

Ongoing Validation:

  • Test on diverse political figures weekly
  • Compare against fact-checkers (Politifact, Snopes, AP)
  • Cross-validate with multiple AI models
  • Incorporate user feedback for improvements

Use Cases

1. Social Media Analysis

> audit @username
Analyze any Twitter/X profile for bias patterns

2. News Article Verification

> audit [paste article text]
> citations [paste specific claims]
Detect framing bias and verify factual claims

3. AI Response Auditing

> benchmark [paste responses from Claude, GPT, Gemini]
Compare how different AIs frame the same topic

4. Self-Auditing

> audit [your own writing]
Check your own bias and logical consistency

5. Research & Academia

> baseline gpt-5
> audit [test responses]
Track AI model behavior changes over time

Contributing

We welcome contributions! Priority areas:

High-Impact Contributions

  1. Marker Expansion - Add new bias detection patterns
  2. Cross-Cultural Adaptation - Adapt for non-US politics
  3. Language Support - Expand beyond English
  4. Validation Testing - Test on more public figures
  5. Platform Optimization - Improve cold-start success rates

Development Priorities

  • CLI tool (Python package)
  • Web interface (bossman.ai)
  • Browser extension
  • API for developers
  • Real-time Twitter/X integration

See CONTRIBUTING.md for guidelines.


Try It Now

Pick your platform and get started:

🚀 Grok Users: Copy /specs/COLD START.txt
🚀 Claude Users: Copy /specs/Claude minimal.txt
🚀 GPT-5 Users: Copy /specs/GPT5 buster.txt

Then audit someone polarizing:

> audit @hasanabi
> audit @benshapiro  
> audit @megynkelly

Share your results:
Tag us with your findings. Let's see who's actually biased. 🎯


Citation

If using BOSSMAN in research:

@software{bossman2026,
  author = {DRockzos, Kyle},
  title = {BOSSMAN Framework v4.1: Bias and Coherence Detection},
  year = {2026},
  url = {https://github.com/[your-username]/bossman}
}

License

MIT License - See LICENSE file for details


Remember: BOSSMAN doesn't tell you what to think.
It shows you how people are trying to make you think.

Use responsibly. Question everything. Including this. 🎯