Defense-in-Depth for Autonomous Agents

Abstract

Large Language Models (LLMs) have demonstrated exceptional capability in generative tasks but struggle with the strict distinctness required for financial due diligence. The probabilistic nature of standard transformer models often leads to "hallucinations"—plausible but factually incorrect assertions—which pose unacceptable risks in investment decision-making.

This paper introduces the architecture behind Gamut Agent, an autonomous deal flow engine designed around a "Defense-in-Depth" methodology. Unlike standard RAG (Retrieval-Augmented Generation) systems, Gamut employs a proprietary Hallucination Firewall™ and a Constraint Verification Engine to mechanically validate data before inference occurs. The result is a system that prioritizes "False Negatives" (rejecting a deal) over "False Positives" (hallucinating a fit), ensuring 100% grounded analysis.

1. Introduction: The Stochastic Problem in Finance

In venture capital and private equity, the cost of a "False Positive" is high. If an AI agent mistakenly identifies a service agency as a SaaS platform, or hallucinates a San Francisco headquarters for a remote company, it wastes human capital and erodes trust.

Standard agentic workflows rely on "Chain of Thought" (CoT) prompting to reason through data. However, CoT is still probabilistic. If the underlying data is missing, an LLM will often "fill in the blanks" based on its pre-training weights rather than admitting ignorance.

Gamut Agent solves this by inverting the workflow. Instead of asking, "What do you think of this company?", the system asks, "Do we have sufficient evidence to form an opinion?" If the answer is no, the Zero-Evidence Circuit Breaker triggers, terminating the process before costs are incurred.

2. System Architecture

The Gamut architecture is defined by three distinct logic gates that govern the flow of information from raw discovery to final verdict.

2.1. The Hallucination Firewall™ (Pre-Inference)

Before any expensive LLM call is made, the target entity undergoes a mechanical verification process. This layer operates deterministically (non-AI).

• DNS & HTTP Verification: The system verifies the existence of the digital footprint.
• Anti-Ghosting Logic: If a domain redirects to a parking page, returns a 404, or lacks valid MX (Mail Exchange) records, the target is classified as a "Ghost."
• Outcome: The system discards the lead immediately. No inference cost is incurred.

2.2. The Zero-Evidence Circuit Breaker

Once a target passes the firewall, the system attempts to harvest data via multi-modal scraping (HTML text, meta tags, and third-party enrichment APIs). The Circuit Breaker evaluates the density of this retrieved data against a minimum viable threshold ($T_{min}$).

This prevents the common failure mode where an Agent reads an empty page and hallucinates a generic business description. If the data payload is empty, Gamut returns a Match Score: 0 with the reason code INSUFFICIENT_DATA.

2.3. Task-Aware Inference Routing

Gamut utilizes a heterogeneous model router to optimize for cost, latency, and reasoning capability.

• Constraint: Geography: Data sensitive to sovereignty (e.g., EU GDPR or China Firewall) is routed to region-compliant local models.
• Constraint: Reasoning: High-complexity tasks (e.g., verifying a "B2B vs. B2C" thesis) are routed to high-parameter Chain-of-Thought models.
• Constraint: Cost: Simple extraction tasks are routed to flash models to minimize overhead.

3. The Constraint Verification Engine

The core innovation of Gamut Agent is the Constraint Verification Engine, which replaces standard "semantic search" with "thesis validation."

3.1. Thesis vs. Query

A "Query" searches for similarity (e.g., "Find AI companies"). A "Thesis" imposes strict constraints (e.g., "Must be Series A AND Location == San Francisco").

3.2. Conflict Resolution & Grounding

The engine ingests conflicting signals from multiple sources to form a synthesized truth.

• Signal A (Enrichment): "Employee Count: 5,000"
• Signal B (Investment Thesis): "Target < 500 Employees"
• Resolution: The engine detects the conflict and issues a NO_FIT verdict, citing Signal A as the disqualifying evidence.

Case Study: The "OpenAI" Stress Test

During internal validation, the engine was tasked with finding "Early-stage AI startups in San Francisco."

• Target: OpenAI.
• LLM Bias: Standard models inherently "know" OpenAI is in SF and might recommend it based on relevance.
• Gamut Behavior: The system ingested the employee count (5,000+) and correctly identified it as violating the "Early-stage" constraint. It issued a REJECT verdict despite the high semantic relevance.

4. Synthetic Evidence Generation

In scenarios where structured data (e.g., industry classification) is missing, but unstructured data (e.g., raw HTML) is rich, Gamut employs a Synthetic Description Generator.

The system dynamically constructs a "Profile" by extracting key phrases from the raw scrape (e.g., "We are a team of 50..." or "Our API documentation..."). This allows the Deep Dive Analyst to proceed with analysis even when traditional database fields are null, ensuring valid startups are not missed due to database gaps.

5. Economic Implications

The Defense-in-Depth architecture fundamentally changes the unit economics of automated due diligence.

By discarding ~40% of noise at the deterministic layer, Gamut reduces the blended cost of due diligence while increasing the reliability of the final output.

6. Conclusion

Gamut Agent represents a shift from Probabilistic Search to Deterministic Verification. By wrapping Large Language Models in rigid, logical safety layers, we enable investors to deploy autonomous agents with confidence. The system does not just find data; it audits it, providing a transparent, verifiable audit trail for every decision.