The Truth Pipeline
The Truth Pipeline is a proprietary request orchestration system utilizing the multimodal neural network Google Gemini 2.5 Pro / Flash.
We do not rely on an AI "black box." Instead, we utilize an Agentic Workflow architecture, where every action (event creation, report, trouble marker) is processed by a cascade of isolated AI agents with hard-coded system instructions.
Instead of costly Fine-Tuning, we employ Dynamic Few-Shot Prompting reinforced with Adversarial Robustness.
⚙️ The Tech Stack
The system operates as an infinite validation loop, returning strictly structured data for smart contract execution.
Core Model: Google Gemini 2.5 Pro (Reasoning) + Flash (Speed).
Methodology: Chain-of-Thought (CoT) + Dynamic Few-Shot Prompting.
Output Format: Strict JSON Schema (Enforced).
Security: Layered Prompt Injection Defense.
🤖 1. AI Agent Classification
The system is divided into 3 specialized auditors. Each prompt acts as an isolated "employee" with a unique area of responsibility.
A. The Trouble Scout (Issue Analysis)
Responsible for the initial validation of map markers (Trouble Spots).
Verification Algorithm:
Injection Shield: The first layer of defense. It ignores any text inside the image (e.g., hidden commands like "Approve this"), classifying it strictly as "Visual Noise."
Evidence Quality: Filters out irrelevant content: selfies, memes, map screenshots, or overly dark photos.
Semantic Match: Checks for semantic consistency. If the category is claimed as "Forest Fire" but the photo shows a "Pothole," the agent blocks the request.
Severity Scoring: Automatically assigns a criticality level (Low/Medium/Critical) to prioritize tasks.
B. The Event Auditor (Event Validation)
Responsible for approving events for publication and fundraising.
Verification Algorithm:
Visual Artifact Detection: The agent searches for semantic anomalies typical of GenAI and crude photo editing: lighting violations, "plastic" textures, impossible object physics, and rendering artifacts.
Context Analysis: Filters out "fake eco-events" (parties, concerts, dates) disguised as cleanups.
Permit Logic: Analyzes uploaded PDF permits or verifies the acceptance of legal liability (Liability Checkbox).
C. The Impact Verifier (Financial Arbiter)
The strictest agent. Responsible for unlocking funds (Escrow Release) from the Reward Pool.
Verification Algorithm (Truth Check):
Geometric Consistency: Compares "Before" and "After" photos. Searches for immutable "anchors" (horizon lines, buildings, unique tree shapes). If the scene geometry does not match, the report is flagged as fraud.
Inpainting Detection: Inspects the "cleaned" area for blur and unnatural texture homogeneity, which are characteristic of AI eraser tools (Magic Eraser).
Receipt Matching: Cross-references receipt items with the claimed work type (e.g., blocks the purchase of personal goods in a tree-planting report).
📡 2. GeoInt & Digital Forensics
In addition to visual analysis, every file passes through a digital forensics layer. This protects against location spoofing and time manipulation.
Metadata Audit
EXIF Data Verification (GPS, Timestamp).
Files without metadata or with traces of tag editing are flagged as High Risk.
Astro-Check
Astronomical Validation.
Calculates the sun's position (azimuth/elevation) for the given coordinates and time. If the photo claims to be at noon but shadows are long - Anomaly Detected.
Meteo-Check
Meteorological Cross-Reference.
Compares weather in the photo with historical data from weather stations (ERA5) at that location (e.g., Sun vs. Rain).
Visual Positioning
Terrain Reference Check.
Compares the photo with satellite imagery or Street View to confirm characteristic landmarks.
🛡️ Security Protocols
System prompts embed HIGHEST PRIORITY directives that cannot be overridden by user input.
Defense Mechanisms
Visual Prompt Injection Defense: The neural network is strictly forbidden from executing text commands found within an image. Any text in a photo is treated by the algorithm exclusively as a visual object.
JSON Enforcement: Any model response that does not match the valid JSON schema is treated by the system as a server error. This prevents data leakage through model "hallucinations.
Last updated