MONNA Semantic Firewall Studio

Project: MONNA Semantic Firewall Studio

Author's Claim: A developer tool for controlling semantic boundaries of ontology-grounded AI agents, using a deterministic semantic firewall to enforce policy-based access control over knowledge graphs.

What Changed: The project is described as a hackathon submission that builds on the idea of separating language understanding from security authority in AI agent systems. It introduces a framework for validating natural-language questions against structured ontologies.

Single Most Important Open Question: Does the author’s self-reported architecture and approach to validation have sufficient grounding to be viable beyond a proof-of-concept, or does it rely on untested assumptions about model behavior and graph alignment?

The description states that MONNA Semantic Firewall Studio is a developer tool designed for controlling the semantic boundaries of ontology-grounded AI agents. It aims to enforce policy-based access control over knowledge graphs.

It consists of three components:

• Capability Prober: Discovers supported question and reasoning paths from an ontology.
• Taxonomy Compiler: Validates question patterns, rejecting those requiring missing or ambiguous relationships.
• Semantic Firewall: Evaluates incoming questions and returns structured ADMIT or REFUSE decisions with reasons, classified paths, matched archetypes, and auditability.

The system uses:

• GPT-5.6 and Codex for interpreting natural-language questions into candidate graph paths.
• A deterministic validation engine to check these against the actual ontology and compiled taxonomy.
• Versioned JSON artifacts for ontologies, taxonomies, and firewall decisions.
• A closed-fail mechanism when model output is invalid or ambiguous.

Inference: The product appears to be a conceptual framework, not yet a deployed tool. It is described as being built during a hackathon with early-stage implementation using GPT-5.6 and Codex.

The author positions MONNA Semantic Firewall Studio as:

• A developer tool for AI agents that operate on structured knowledge.
• A solution to the problem of AI agents producing convincing but unauthorized answers.
• A system that enforces policy-based access control, not just keyword filtering.

Key claims:

• The system is built around the principle: “The floor is policy. The ceiling is the graph.”
• It separates language understanding from security authority.
• It uses LLMs to interpret questions, but graph validation makes final decisions.
• It supports auditability and version control.

Inference: This is a conceptual positioning of a tool that could be part of a broader category of AI governance or agent control systems. The author does not claim it is in production, nor does the description suggest any prior development beyond a hackathon prototype.

The description states:

• MONNA Semantic Firewall Studio is designed for developers working with ontology-grounded AI agents.
• It targets users who need to control semantic boundaries and ensure that agent responses are grounded in authorized knowledge sources.

Inference: The target customer appears to be technical developers or AI engineers, likely within organizations using or building AI agents with structured knowledge bases (e.g., RAG systems, knowledge graphs). However, no specific customer segments or personas are named.

The description does not contain any information about:

• Pricing
• Revenue model
• Monetization strategy
• Customer acquisition plans

Not evidenced

The system is described as being built with:

• GPT-5.6
• OpenAI Codex
• React and TypeScript
• JSON-based artifacts for ontologies, taxonomies, and firewall decisions

It includes:

• A deterministic validation engine
• Versioned artifacts
• Fail-closed behavior in ambiguous or invalid cases
• Audit logs and reason codes
• A visual interface for understanding why a question was admitted or refused

Inference: The technical approach is described as hybrid, combining LLM interpretation with deterministic graph validation. It suggests an architecture that prioritizes inspectability, auditability, and safety over automation.

The description states:

• This is a hackathon submission.
• The system is currently in early conceptual exploration.
• It includes a runnable Build Week implementation, but no production deployment or customer feedback is mentioned.

Not evidenced

The description does not mention:

• Competitors
• Market positioning relative to other AI governance tools
• Existing solutions in the space of agent control, knowledge graph validation, or semantic firewalls

Not evidenced

• Unproven assumptions: The system relies on GPT-5.6 to interpret questions and propose candidate paths, but it is unclear how well this model will generalize across different ontologies or handle adversarial inputs.
• Limited validation scope: The project is described as a hackathon prototype with no mention of testing against real-world data or edge cases.
• No evidence of scalability or integration: There is no indication that the system can be integrated into existing AI agent workflows or scaled for enterprise use.
• Single-person team: The team size is listed as one, which may limit development velocity and depth.

1. What specific types of ontologies or knowledge graphs does this tool support?
2. How does the system handle cases where the model’s interpretation is ambiguous but the graph allows a path?
3. Has the validation engine been tested against adversarial inputs or edge cases?
4. What are the limitations of the current JSON-based artifact format in terms of interoperability with other tools (e.g., RDF, OWL)?
5. How does the system handle version mismatches between ontologies and taxonomies?
6. Are there any plans for integrating with CI/CD pipelines or existing agent frameworks?