Zero Mental Math: An Anti-Hallucination Architecture for LLM-Driven Analysis

Zero Mental Math: An Anti-Hallucination Architecture for LLM-Driven Analysis

A six-layer system for achieving 100% accurate numerical reporting from Large Language Models. The architecture shifts all computation to deterministic Python code, reducing LLMs to “citation copy machines” with zero hallucination risk.

Why This Matters

LLMs are fundamentally pattern matchers, not calculators. When asked to analyze numerical data, they generate “plausible-looking” numbers based on statistical patterns in training data—not deterministic computation. This leads to catastrophic failures in domains like financial reporting, where even a 0.1% error in win rate calculations can result in millions of dollars in losses. Research shows LLMs fail reliably at multi-digit arithmetic (Nogueira et al., 2021), making this an architectural limitation, not a bug.

Key Insights

• “100% accuracy on end-to-end integration tests, 2025”: The system’s MCP server guarantees deterministic results through pre-calculated metrics.
• “Citations over generation for financial reporting”: By forcing LLMs to copy pre-formatted citations, the architecture eliminates paraphrasing errors.
• “ChromaDB for RAG in financial systems”: Semantic grounding via knowledge bases prevents entity resolution hallucinations.

Working Example

# ❌ BAD: Raw data requires LLM to calculate
get_mt5_history_deals() → [deal1, deal2, deal3, ...]

# ✅ GOOD: Pre-calculated metrics
get_mt5_position_history() → {
    "summary": {
        "total_positions": 29,  # Server counted
        "win_rate": 65.52,      # Server calculated: (19/29)*100
        "profit_factor": 2.34,  # Server calculated: sum(wins)/abs(sum(losses))
        "expectancy": 42.57     # Server calculated: total_pl/total_positions
    }
}

{
    "_accuracy_report": {
        "checksum": "A7B3C2D1",
        "checksum_input": "29|19|10|65.52|1234.56|85.25|-42.15|2.34|42.57",
        "metrics": [
            {
                "path": "summary.win_rate",
                "value": 65.52,
                "citation": "Win rate: 65.52% [Source: get_mt5_position_history.summary.win_rate]"
            }
        ]
    }
}

# Template-based output formatting
TEMPLATE = """## Performance Analysis (Confidence: {confidence.score})
### Overview
{citation:summary.total_positions}
{citation:summary.win_rate}
{citation:summary.profit_factor}
[Verified: {checksum}]"""

Practical Applications

• Use Case: Financial reporting systems using Python-based MCP servers to pre-calculate trading metrics.
• Pitfall: Relying on LLMs for arithmetic without server-side checks, leading to hallucinations like “16th trade when only 15 exist.”

References:

• https://arxiv.org/abs/2302.04761
• https://arxiv.org/abs/2210.03629
• https://arxiv.org/abs/2203.11171
• https://arxiv.org/abs/2212.08073
• https://arxiv.org/abs/2303.17651
• https://arxiv.org/abs/2211.10435
• https://arxiv.org/abs/2102.13019
• https://dev.to/nodefiend/trust-the-server-not-the-llm-a-deterministic-approach-to-llm-accuracy-20ag

---