Systèmes LLM à vote : leçons de la NASA pour la fiabilité de l’IA

Executive Summary

Large Language Models (LLMs) have revolutionized business automation, but their tendency to "hallucinate" false information poses serious risks in mission-critical applications. This paper presents a voting-based approach inspired by NASA's spacecraft redundancy systems, where multiple LLMs process the same query and vote on results. This methodology dramatically improves accuracy while automatically flagging potential errors for human review.

1. Introduction: The Cost of Being Wrong

In government procurement, marine hardware specification, and defence contracting, accuracy isn't just important — it's legally and financially critical. A single error can result in:

• Financial Loss: Wrong quantities or specifications in DND contracts can cost thousands in returns, restocking, or contract penalties
• Safety Issues: Incorrect material grades for anchor chains or marine hardware can lead to equipment failure
• Legal Violations: Export control misclassification (ITAR, EAR) can result in severe penalties
• Reputation Damage: Delivering incorrect specifications to government clients damages long-term relationships
• Operational Delays: Wrong parts halt production lines and delay mission-critical operations

2. The NASA Lesson: Redundancy Over Perfection

From Single Points of Failure to Voting Systems

In the early days of spaceflight, NASA faced an impossible challenge: computers weren't reliable enough to trust with human lives, yet missions were too complex to fly without them. The solution wasn't to build one perfect computer — it was to build multiple computers that checked each other's work.

The Core Insight

NASA's breakthrough wasn't building perfect systems — it was building systems that could detect their own failures through consensus. When computers agreed, confidence was high. When they disagreed, automatic alerts triggered human intervention.

This same principle applies perfectly to AI/LLM deployment in business-critical applications.

3. The LLM Hallucination Problem

What Are Hallucinations?

LLM "hallucinations" occur when AI systems generate plausible-sounding but factually incorrect information. Unlike random errors, hallucinations are particularly dangerous because they appear confident and well-formatted, making them difficult to spot without domain expertise.

Common Hallucination Types in Business Applications

• Number Transposition: Reading "M2700512" as "M2750012"
• Specification Confusion: Confusing Grade 3 with Grade 4 steel specifications
• Context Misinterpretation: Misunderstanding technical abbreviations or industry jargon
• Fabricated Details: Adding plausible but nonexistent information to fill gaps
• Unit Conversion Errors: Incorrectly converting between metric and imperial measurements

Input: "Supply 100 ea. Anode, Corrosion Preventive, NSN 5340-15-017-5426,
        P/N M2700512, NCAGE A5900, Manufacturer: ONDA SP (Italy)"

Single LLM Output:
- NSN: 5340-15-017-5426 ✓
- Part Number: M2750012 ✗ (transposed digits)
- Quantity: 100 ✓
- Cage Code: A5900 ✓

Result: Wrong part ordered, $8,500 loss + procurement delay

4. Voting LLM Systems: The Modern Application

The Architecture

User Query: "Extract NSN, P/N, and quantity from solicitation"
                    ↓
    ┌───────────┼───────────┬───────────┐
    │           │           │           │
┌───▼───┐   ┌───▼───┐   ┌───▼───┐   ┌───▼───┐
│ LLM 1 │   │ LLM 2 │   │ LLM 3 │   │ LLM 4 │ (optional)
│Claude │   │ GPT-4 │   │Gemini │   │Claude │
└───┬───┘   └───┬───┘   └───┬───┘   └───┬───┘
    │           │           │           │
    └───────────┼───────────┴───────────┘
                ↓
        Voting Engine
     (Compare Responses)
                ↓
        ┌───────┴────────┐
        │  All Agree?    │
        └───┬────────┬───┘
        YES │        │ NO
            ↓        ↓
       Automatic  Flag for
       Approval   Human Review

Implementation Strategies

5. Practical Implementation at Dibblee Industries

Use Case 1: DND Contract Processing

Processing Results (47-page DND solicitation for compressor units):

LLM 1 (Claude):  NSN 4310-01-234-5678, P/N CP-8800-12, Qty: 24, Date: 2026-06-15
LLM 2 (GPT-4):   NSN 4310-01-234-5678, P/N CP-8800-12, Qty: 24, Date: 2026-06-15
LLM 3 (Gemini):  NSN 4310-01-234-5678, P/N CP-8800-12, Qty: 24, Date: 2026-06-15

VOTING RESULT: ✓ CONSENSUS — Automatic approval
Confidence: 100% | Human review: NOT REQUIRED

LLM 1 (Claude):  Material: "316 Stainless Steel"
LLM 2 (GPT-4):   Material: "316 Stainless Steel"
LLM 3 (Gemini):  Material: "304 Stainless Steel"

VOTING RESULT: ⚠ DISAGREEMENT DETECTED
Confidence: 67% | Human review: REQUIRED
Flag: Material specification ambiguous in source document

Use Case 2: Technical Drawing Interpretation

When processing technical drawings for supplier distribution:

• Each LLM analyzes the drawing and extracts dimensions, tolerances, and material callouts
• Voting consensus validates measurements and specifications
• Disagreements flag ambiguous or poorly-scanned drawings for human review
• Result: Reduced supplier queries and improved first-time-right manufacturing

Use Case 3: Export Control Classification

Product: Naval anchor chain, Grade 4, 2-inch links

LLM 1: "ITAR controlled — military application"
LLM 2: "EAR99 — commercial item"
LLM 3: "ITAR controlled — military application"

VOTING RESULT: ⚠ NO CONSENSUS (2-1 split)
Action: MANDATORY human review by compliance officer
Risk Level: CRITICAL — legal penalties possible

Outcome: Human expert determines item is ITAR-controlled when sold to DND for military vessels, but EAR99 for commercial marine use. The disagreement correctly identified the complexity requiring expert review.

6. Cost-Benefit Analysis

The Math Behind Voting Systems

Real Cost Comparison: Monthly DND Processing

7. Implementation Roadmap

8. Technical Architecture Details

Voting Engine Pseudocode

async function votingLLMQuery(prompt, field_criticality) {
    const models = field_criticality === 'critical'
        ? ['claude-3-5-sonnet', 'gpt-4', 'gemini-pro', 'claude-3-5-sonnet', 'gpt-4']
        : ['claude-3-5-sonnet', 'gpt-4', 'gemini-pro'];

    const responses = await Promise.all(
        models.map(model => callLLM(model, prompt))
    );

    const parsed = responses.map(r => parseStructuredOutput(r));

    const results = {};
    for (const field of Object.keys(parsed[0])) {
        const votes = parsed.map(p => p[field]);
        const consensus = findConsensus(votes);

        results[field] = {
            value: consensus.value,
            confidence: consensus.percentage,
            requires_review: consensus.percentage < 100,
            all_votes: votes
        };
    }

    return results;
}

function findConsensus(votes) {
    const frequency = {};
    votes.forEach(v => {
        frequency[v] = (frequency[v] || 0) + 1;
    });

    const max = Math.max(...Object.values(frequency));
    const winner = Object.keys(frequency).find(k => frequency[k] === max);

    return {
        value: winner,
        percentage: (max / votes.length) * 100,
        vote_breakdown: frequency
    };
}

Integration with Existing Systems

• MCP Server Integration: Voting engine exposed as MCP tool for Claude to call
• Database Layer: Store all votes, disagreements, and resolutions in PostgreSQL for audit trail
• Xero Integration: Validated contract data flows directly into invoice generation
• Human Review Dashboard: Web interface showing flagged items with side-by-side LLM responses
• Monitoring: Track consensus rates, processing costs, and accuracy metrics over time

9. Measuring Success

Key Performance Indicators

10. Risk Mitigation and Fallback Strategies

11. Competitive Advantage and Market Positioning

Implementing voting LLM systems positions Dibblee Industries as a technology leader in a traditionally conservative industry:

• Client Confidence: Government clients value reliability above all. 99%+ accuracy with automated hallucination detection builds trust.
• Faster Response Times: Same-day bid responses instead of multi-day turnarounds.
• Scalability: Handle 10× contract volume without proportional staff increases.
• Audit Trail: Complete record of all AI decisions and human reviews satisfies government compliance.
• Knowledge Transfer: System captures expert decisions, reducing dependency on individual staff.
• Service Offering: Potential to offer AI-validated procurement services to other distributors.

12. Conclusion: The Path Forward

The voting LLM approach represents a paradigm shift in how we think about AI reliability. Rather than pursuing the impossible goal of a single perfect AI, we embrace the NASA lesson: consensus reveals truth, and disagreement reveals uncertainty.

For Dibblee Industries, this isn't just about automation — it's about building a fundamentally more reliable procurement process than was ever possible with purely human operations.

Next Steps

1. Week 1: Select pilot project (recommend: DND contract extraction)
2. Week 2: Implement 3-LLM voting proof-of-concept
3. Week 3–4: Process 50 test cases and measure results
4. Week 5–6: Build production voting engine and review dashboard
5. Week 7+: Scale to additional use cases and optimize