Physics-Based AI Integrity System
for Japanese Property Insurance

Protecting Japan's culture of high trust from generative AI fraud. We reverse-engineer the physical laws of photographed scenes to secure automated digital workflows.

The Threat

Japan's working-age population will drop by 60% by 2065. Property insurers must rely on full digital automation. Scammers worldwide now generate highly realistic, fake photographs of flooded rooms and structural damage to steal insurance funds. Standard pixel-based detectors fail against modern algorithms.

The TRISOL Solution

We map 3D depth to verify perspective rules and shadow consistency across multiple objects. Built entirely on safe open-source licenses (LGPL v3.0, YOLO-NAS Open Weights), allowing Japanese corporations to deploy our software completely on-premise with zero legal risks for their private code.

Phase 1: Real-World Calibration

Before detecting fakes, TRISOL establishes a baseline of physical reality. We utilize the advanced YOLO-NAS architecture (by Deci.ai) for precise object and shadow detection, mapping physical anchors within the room.

Baseline: Real Environment (Score: 90/100)

Reading the Report: The Master Report below demonstrates TRISOL's multi-view analysis. The colored bounding boxes show YOLO-NAS identifying distinct physical objects. The connecting perspective vectors (cyan lines) prove that all objects exist in the same 3D space and obey the same gravitational and lighting laws. The high score confirms the scene's organic spatial consistency.

Original Photo Inputs (Click to zoom)

Real 1 Real 2

TRISOL Master Report (Click to view full resolution)

Real Photo Analysis

Phase 2: Defeating SOTA AI Generators

State-of-the-Art (SOTA) models generate flawless textures that easily bypass standard pixel-analysis tools. However, they lack a fundamental understanding of 3D physics. Defeating these specific industry-leading models proves that TRISOL's geometric reverse-engineering is future-proof against the highest levels of digital fraud.

Adversarial Attack: FLUX.2 Pro (Score: 66/100)

The Threat: Developed by Black Forest Labs, FLUX.2 Pro is currently recognized as the leading model for hyper-realistic architectural generation and strict prompt adherence. It is a prime tool for sophisticated insurance fraud.

Reading the Report: While visually flawless to the human eye, TRISOL mathematically exposes the forgery. The red and yellow vectors in the report highlight a critical 23.6-degree geometry error in the dining table's perspective when mapped across two different generated "angles". The physics engine flags the claim for manual review.

Generated Inputs: FLUX.2 Pro (Click to zoom)

Flux 1 Flux 2

TRISOL Master Report (Click to view full resolution)

FLUX.2 Pro Analysis

Adversarial Attack: Gemini Advanced (Score: 40/100)

The Threat: Google's Gemini multimodal ecosystem represents top-tier consumer AI. Its ability to iterate on images makes it dangerous for generating multiple "evidence" photos of a fake scene.

Reading the Report: This dashboard shows a total collapse of spatial mapping. The AI attempted to generate 4 different perspectives of the same room. TRISOL's object detection maps severe structural shifting: furniture changes scale and coordinates relative to the walls between shots. The system assigns a critical failure score of 40/100, blocking the automated claim.

Generated Inputs: Gemini Advanced (Click to zoom)

Gemini 1 Gemini 2 Gemini 3 Gemini 4

TRISOL Master Report (Click to view full resolution)

Gemini Analysis

The Scientific & Engineering Core

Dr. Oleksii Kofanov

Co-Founder & Project Lead

Ph.D. in Economics & Ph.D. in Engineering Sciences. Manager of the EU HORIZON grant EU-TRAINS. Cabinet of Ministers of Ukraine Scholarship for Young Scientists 2024-2026.

LinkedIn ORCID

Prof. Olena Kofanova

Co-Founder & Head of R&D

Ph.D. in Chemistry & Doctor of Pedagogical Sciences. Expert Reviewer for the IPCC. Winner of the Innovative Ideas Contest 2023 (IDEA-East-Hub, Germany).

LinkedIn ORCID