AI Model Card

LLM Auditor Agent (v.2) Gemini 3.5 Flash

1. Model Overview

Model Name: LLM Auditor Agent
Underlying Generative Engine: Gemini 3.5 Flash (gemini-3.5-flash)
Client Framework: Unified google-genai SDK (v2.0+)
Model Type: Multi-Agent Verification & Revision System (Critic & Reviser Pipeline)
Deployment Platform: Google Cloud Run (us-central1) & Vertex AI Engine
Data Format: Accepts User Question + LLM Output payload; returns verified JSON (audit_report and revised_llm_answer).

Core Function: Automated, verifiable content auditing, hallucination detection, and schema-enforced corrective revision.
Intended Usage: Enterprise compliance, legal, financial, and technical auditing to prevent reputational damage and factual errors in automated LLM workflows.
Performance Baseline: Sub-second audit latencies with zero-trust context grounding.

The LLM Auditor Agent operates as a dedicated security and verification layer built atop Gemini 3.5 Flash:

Vertex AI RAG Grounding: Evaluates responses against secure, customer-provided knowledge bases (Vertex AI Search / Discovery Engine) to eliminate ungrounded foundation model assertions.
Multi-Agent Redundancy:
- Critic Agent: Conducts strict factual checking and risk analysis against retrieved context.
- Reviser Agent: Generates corrected, structured JSON outputs to ensure unverified data is never released.
Schema Sanitization: Enforces rigid JSON schema contracts to prevent unparsed or malformed model responses.
Full-Stack Observability: Native integration with Google Cloud Operations Suite (Cloud Logging, Cloud Monitoring, and Vertex AI telemetry) to track request latencies, token consumption, and safety metrics in real-time.

Hallucination & Factual Accuracy: Base generative models may hallucinate facts or rely on stale knowledge.
PFZ Mitigation: RAG Grounding Layer forces real-time validation against customer-provided data stores, neutralizing baseline knowledge cutoffs.
Model Misinterpretation: System may misinterpret user intent or complex document structures.
PFZ Mitigation: Critic-to-Reviser Chain dual-stage verification ensures uncorrected outputs are caught and revised prior to presentation.
Data Reliance: Audit accuracy is dependent on the quality of ingested customer data.
PFZ Mitigation: Restricted Scope confines audit boundaries strictly to authorized customer data stores.
Prompt Injection & Overreliance: Adversarial input queries or ungrounded prompts.
PFZ Mitigation: Governed by Google's Generative AI Safety Settings and PFZ Zero-Trust execution boundaries.

Google Cloud Partner Status: Deployed natively on Google Cloud infrastructure (Project pfz-llm-auditor-agent-prod).
Ethics & Transparency: Fully compliant with the PFZ AI Governance Framework, Proactive Transparency Policy, and global Anti-Bribery standards. All outputs carry explainable reasoning paths.