TOON Protocol: Token Optimization and FDIA Gating in Delentia OS

Building production LLM applications for the enterprise requires a careful balance between safety and resource consumption. In Delentia OS, we resolve this tension using two core techniques: TOON Protocol for token optimization, and FDIA Gating for safety consensus.

Together, these mechanisms allow us to compress token overhead by 38.4% while keeping hallucination rates below 0.3%.

1. TOON Protocol: Reducing Token Overhead by 38.4%

In multi-agent systems, models spend a significant portion of their token budget discussing routing, state representation, and system instructions.

The Token-Optimized Orchestration Network (TOON) Protocol implements dynamic payload compression for inter-agent communications. Rather than passing verbose natural language structures between specialized LLMs, TOON Protocol abstracts state updates into highly compressed symbolic structures.

Key Optimization Pillars:

1. State Deltas Only: Instead of passing the entire prompt context, agents transmit only state deltas.
2. Grammar Constraints: Leveraging JSON schemas and strict compiler-guided formats directly on the LLM logit biases to prevent token waste in formatting.
3. Intent Caching: Reusing computed intent embeddings to bypass repetitive system prompt ingestion.

In benchmark tests across 10,000 corporate documents, the TOON Protocol achieved an average of 38.4% token reduction with zero loss in semantic accuracy or execution quality.

2. Gating Intents via the FDIA Equation

To ensure data sovereignty and rule-based compliance (especially under frameworks like Thailand's PDPA), Delentia OS filters every system request using the FDIA Gating Equation:

$$F = D^I \times A$$

Where:

• F represents the future outcomes and output security score.
• D represents the encoded corporate data.
• I represents the system's intent amplification vector.
• A represents the Architect (Human-in-the-Loop veto power).

By feeding user inputs through this gate, the operating system can verify whether the system's generated intents are aligned with organizational boundaries before initiating downstream multi-LLM consensus. If the Architect value is zero ($A=0$), the execution is immediately vetoed, preventing high-risk actions before they ever reach the models.

3. Real-World Applications

For enterprise workloads requiring offline or air-gapped deployment, minimizing token volume directly translates to lower local GPU RAM requirements. Lower token counts allow smaller specialized models (such as 8B or 14B parameter models running on local Ollama or HuggingFace adapter layers) to process larger window sizes without memory overflow.

If you want to experience these protocols first-hand, join our developer preview. Select your target infrastructure (Docker, Kubernetes, or Air-Gapped) and request access through our terminal CLI to receive the detailed architecture whitepaper.

This article was published by the Delentia Labs Research Team. For inquiries regarding licensing or local deployments, please contact us at founder@delentia.com.