> ## Documentation Index
> Fetch the complete documentation index at: https://hotchilianalyticsllc.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Benchmarks & Performance

> How VerityNgn performs against real-world misinformation

# Benchmarks & Performance

VerityNgn has been evaluated on a **200-claim test set** spanning health supplements, financial advice, cryptocurrency, and viral misinformation.

## Key Metrics

| Metric                        | Result                     |
| :---------------------------- | :------------------------- |
| **Accuracy vs. Ground Truth** | 78% (95% CI: 61-85%)       |
| **Counter-Intel Lift**        | +18% on misleading content |
| **Calibration (Brier Score)** | 0.12                       |
| **Precision (False Claims)**  | 85%                        |

### The "Counter-Intel" Advantage

The counter-intelligence module is the primary differentiator for VerityNgn. By actively searching for debunking videos and detecting promotional press releases, we achieve an **18 percentage point improvement** in accuracy when analyzing misleading content (scams, conspiracy theories).

## Efficiency (v2.0)

With the introduction of **Intelligent Video Segmentation**, VerityNgn v2.0 significantly reduces costs and processing time by maximizing the 1M token context window of Gemini 2.5 Flash.

| Metric                       | v1.0 (Fixed) | v2.0 (Intelligent) | Improvement         |
| :--------------------------- | :----------- | :----------------- | :------------------ |
| **API Calls** (33-min video) | 7            | 1                  | **86% Reduction**   |
| **Processing Time**          | 56-84 min    | 8-12 min           | **\~7x Faster**     |
| **Context Utilization**      | \~3%         | \~58%              | **19x Improvement** |

## Token Economics

Processing a typical video costs between **$0.50 – $2.00** depending on:

1. Video length (impacts multimodal input tokens).
2. Number of claims extracted (impacts output tokens).
3. Search depth (Google Search API costs).

## Error Calibration

We avoid binary "True/False" labels. Instead, we use a calibrated three-state distribution:
`P(TRUE) + P(FALSE) + P(UNCERTAIN) = 1.0`

This ensures that when the system is unsure (e.g., a "Likely False" verdict with 40% Uncertainty), it signals the need for human review rather than forcing a potentially incorrect binary choice.

## CRAAP Analysis

The system performs a deep credibility analysis, scoring content on the industry-standard **CRAAP Analysis** framework:

* **Currency**: When was the information published?
* **Relevance**: Is it relevant to the video's main message?
* **Authority**: What credentials or expertise is claimed?
* **Accuracy**: Can it be verified with external sources?
* **Purpose**: Is it promotional, educational, or persuasive?