The Prompt Paradox: Moving to Deterministic AI Scoring

The Prompt Paradox

In agentic systems, instructing an LLM to be simultaneously "genuine" and to "produce a rigid integer score" creates what we call the Prompt Paradox. The model attempts to fulfill both constraints by hyper-fixating on minor deviations to justify its numerical output, leading to sporadic and non-deterministic scoring volatility.

For enterprise risk management, a contract risk score cannot fluctuate wildly based on the temperature of a language model. It must be mathematically stable.

The Deterministic Fix

To resolve the Prompt Paradox, we decoupled the analytical reasoning from the final scoring calculation.

1. 1.Extraction Only: The Mike AI agent is now strictly tasked with extracting and classifying risk vectors (High, Medium, Low) based on the enterprise playbook. It no longer attempts to guess a final score out of 100.
2. 2.Python-Based Calculation: After the LLM returns the JSON payload, our deterministic engine calculates the final score natively: Base Score (100) - (High Risks * 20) - (Medium Risks * 10).

Unshakeable Stability

This architectural pivot ensures that identical risk vectors always produce identical risk scores, restoring absolute trust in the Intelligence Hub while allowing the LLM to focus purely on semantic analysis.