KDG Technical Paper
What is the 4-Layer Knowledge Architecture
This paper describes the technical structure of KDG's core 4-layer knowledge architecture. It introduces the 'Text-Driven Principle' — which keeps domain knowledge as human-readable text rather than locking it into black boxes — and explains the role of each of the four knowledge layers (Original / Formal / Insight / Meta), comparing them with common AI approaches.
3.Detailed Implementation of the 4 Layers
This chapter describes the role of each of the four layers and the typical format of their Markdown descriptions. All layers are managed and edited as text within the KBMS (Knowledge Base Management System) of KDG STUDIO.
3-1Original Layer — Source Retention
The Original layer stores source data owned by the organization as-is. Regulations, manuals, daily reports, past cases, specifications, meeting minutes — any business document is a target. What matters is that these are held as 'reference data' rather than used as 'training data' for machine learning.
The role of the Original layer is to always keep source documents accessible as the ultimate basis for decisions. No matter what interpretations or structuring later layers apply, returning to the Original layer provides access to first-hand information. This source retention mechanism plays a particularly important role in audit operations where AI conclusions must be verified after the fact, and in business domains where transparency in decision-making is required.
3-2Formal Layer — Structured Data
The Formal layer stores 'structured decision rules' extracted from the Original layer's source text. Since source text alone cannot be used for judgment, the Formal layer organizes it into explicit formats such as rules, procedures, and patterns. The typical composition is a three-part structure of 'Ideal State,' 'Valid Examples,' and 'Invalid Examples.'
### Ideal State (Description of expected state or decision criteria) ### Valid Examples - (Example of correct input or condition) - (Example of correct input or condition) ### Invalid Examples - (Example of incorrect input) → (NG reason template) - (Example of incorrect input) → (NG reason template)
What makes this format excellent is that decision logic and explanation reasons coexist in the same text. Because 'invalid examples' serve directly as NG reason templates, judgment and explanation are structurally aligned. This allows AI to consistently generate both the decision result and its reason simply by reading the Formal layer text. Rule changes are reflected instantly by editing the text.
3-3Insight Layer — Derived Knowledge
The Insight layer stores 'cross-item insights' extracted from relationships between multiple check items and data. It is a layer that holds larger patterns and best practices beyond individual decision rules. For example, insights such as 'specific combinations of conditions typically become risk factors' or 'success patterns common to past similar cases' are handled here.
## Rule: When [condition], then [action] ### Rationale [Supporting evidence from past data or domain expertise] ### Conditions [Explicit application conditions] ### Target Items - [List of items affected by this rule]
Insight layer rules complement the individual rules of the Formal layer. By making explicit the relationships invisible from individual items, they turn organizational experience into reproducible assets. Such insights, which would be buried as 'implicit patterns' in learning-based approaches, are organized in the Insight layer in a form that humans can read and edit.
3-4Meta Layer — Values & Policy
The Meta layer stores organizational review policies, priorities, approval authority, and exception rules. It is the highest-priority layer among the four, capable of overriding decisions from any other layer. The entire decision-making framework — 'how the organization should judge' — is managed as text.
### Mandatory Requirements (If not met, NG by default) - Requirement 1 - Requirement 2 ### Exception Rules (Patterns for conditional approval) - Pattern 1: [condition] → [approval condition] - Pattern 2: [condition] → [approval condition] ### Approval Authority | Risk Level | Approver | |---|---| | Low | [role] | | Medium | [role] | | High | [role] |
The significance of the Meta layer lies in the fact that when organizational policy changes, that change is immediately reflected in all decisions. Learning-based approaches like LoRA require retraining every time policies change, but KDG only requires editing the Meta layer text. The moment organizational decision-making shifts, AI behavior follows — this is the role of the Meta layer.
3-5Integration of the 4 Layers
The four layers do not operate independently. At decision time, they coordinate in the following order.
- 1Meta layer (highest priority): Check whether organizational policy applies an override
- 2Insight layer: Determine whether cross-item patterns apply
- 3Formal layer: Match against individual item rules
- 4Original layer: Reference source documents as needed
Through this hierarchical structure, information abstraction increases step by step from 'individual rules → cross-cutting insights → organizational policy,' and the final decision aligns with the organizational decision-making framework. Because all four layers are managed as human-readable text, tracing the basis for decisions in reverse is also straightforward.
4.Constraints and Conclusion
4-1KDG's Constraints and Tradeoffs
The 4-layer knowledge architecture is not universal. The following constraints must be understood when adopting it.
Knowledge Construction Cost
Building the 4-layer knowledge requires humans with domain expertise. AI assists this knowledge extraction process, but final approval of decision criteria is done by humans. However, if one intends to leverage AI's reasoning capabilities for business judgment, explicitly organizing the knowledge that underlies that reasoning is an inherently unavoidable cost — and KDG is designed to structurally support the value that justifies it: accuracy, reproducibility, and accountability.
Applicable Domain
KDG is optimized for operations requiring 'judgment based on explicit rules.' For routine FAQ responses where decision criteria are straightforward, KDG is overengineered — conventional RAG handles these adequately. Conversely, for complex challenges where defining correct answers is difficult, sufficient knowledge formalization is a prerequisite, making it important to narrow the domain scope and evaluate applicability incrementally.
Growth Curve
The maturity of both the Original layer and the Formal layer determines KDG's judgment accuracy. In the early stages, it is recommended to focus on limited answer patterns and begin operations within a scope where judgments can be made reliably. As operational experience accumulates and both the Original layer's historical data and the Formal layer's judgment rules become enriched, the scope of application can be expanded incrementally — this phased approach is the most practical path for KDG deployment.
4-2Conclusion
This paper has described the technical structure of KDG's core 4-layer knowledge architecture. The essence of this architecture lies in the Text-Driven Principle: organizational domain knowledge is explicitly managed as human-readable, editable Markdown text — not as model weights or vector spaces. This transforms AI execution from inference-based to rule-matching-based, aiming to achieve both deterministic decision results and accountability.
- ▸The Original layer always retains the organization's source data in a referenceable form
- ▸The Formal layer explicitly states rules for each decision item in a three-part structure: 'Ideal State,' 'Valid Examples,' 'Invalid Examples'
- ▸The Insight layer extracts cross-item insights as automatic decision rules
- ▸The Meta layer manages the organizational decision-making framework as highest-priority rules
Critically, this 4-layer structure is a model-independent principle of knowledge management. Accumulated knowledge is never lost when AI models are swapped, and switching to a more capable model automatically yields more advanced Insights from the same knowledge base. KDG's approach — positioning knowledge not as 'material for guessing' but as 'foundation for deterministic execution' — is a design for returning organizational knowledge to the hands of the organization. When the customer's business knowledge exists in a form visible to the customer, this simple principle transforms AI into a trustworthy business tool.