Trust the Server, Not the LLM: A Deterministic Approach to LLM Accuracy

The article says LLMs can hallucinate numbers in trading and financial reporting because they pattern-match instead of compute. Its proposed architecture moves all calculations to deterministic Python, adds citation and checksum tracking, constrains output with templates, grounds entities with RAG, validates drafts with a second LLM, and retries corrections when validation fails.

Reading notes

#

• The problem is an MCP server pulling MT5 trading data while the LLM invents values such as an extra trade count.
• The core claim is that LLMs are not calculators and fail at arithmetic tasks, especially multi-digit or more complex ones.
• Layer 1 makes the MCP server compute totals, win rate, profit factor, and expectancy before the LLM sees the data.
• Layer 2 adds an _accuracy_report with preformatted citations, a CRC32 checksum, and a confidence score.
• Layer 3 uses templates so the response formatter only fills citation slots and does not calculate anything.
• Layer 4 retrieves static facts from ChromaDB so entity mappings and trading rules stay grounded.
• Layer 5 sends the draft to a second LLM for strict verification against source data and checksum rules.
• Layer 6 parses validation errors, applies corrections, and retries up to a fixed limit.
• The anti-aggregation rule requires raw values before averages so trends and changes are visible.
• Final output is expected to show citation tags, a checksum, confidence information, and no approximations.