The Two Sides of the Coin: Understanding Arche AI's Deception Score vs. Confidence Level

Layer 1: The Core Deception Score (The "What")

First, it is important to remember that our system performs multiple, separate analyses that inform each other. The initial and most fundamental analysis is our primary deception model. When you see a segment marked as "Deception Detected," it has already passed this first test with a score greater than 60%.

This primary score is generated by analyzing the spectral data of the voice itself—the non-linguistic, subconscious vocal biomarkers that indicate cognitive load. This is the core "what" of our analysis; it's the raw signal that flags a moment for a closer look.

Layer 2: The Confidence Level (The "Why")

Once a segment is flagged, our Large Language Model (LLM) acts like a detective to analyze the language and context. The confidence level you see is a measure of the LLM's certainty in its own linguistic analysis of that specific segment.

Our AI evaluates the strength of linguistic indicators, the clarity of the surrounding conversation, and the consistency of speech patterns to assign its confidence.

What the Different Levels Mean

Here's a quick guide to interpreting the confidence levels:

• High Confidence (>80%): The AI is very certain about its linguistic interpretation. The context is clear, and the verbal patterns are consistent. This is a strong signal to investigate further.
• Medium Confidence (50-80%): The AI has moderate certainty. There may be some ambiguity in the context or mixed signals in the speech patterns.
• Low Confidence (<50%): The AI is less certain. This could be due to limited context, unclear speech, or conflicting linguistic indicators.

The Most Important Takeaway

You could have a segment with a 90% deception score but only a 60% confidence level.

What does this mean? It tells you that while our primary spectral analysis strongly suggests deception (the 90% score), our LLM is only moderately confident in its own linguistic assessment of why (the 60% score), perhaps due to a lack of surrounding context in the conversation.

This dual-metric system is designed to help you prioritize. A high deception score with low confidence doesn't invalidate the finding; it simply tells you that more human investigation is needed to truly understand the context behind the signal.

By using both the deception score and the confidence level, you can more effectively pinpoint which moments in a conversation deserve the most attention, turning complex analysis into clear, actionable intelligence.