A Machine Learning Infrastructure for Assetizing and Evolving Decision Data — Decision Trace Ledger × GNN Core

■ Introduction

Until now, AI has evolved as a system that handles signals, such as:

• Predicting
• Classifying
• Generating

However, in real-world operations, what truly matters is not:

👉 what was produced,

but rather:

👉 what was decided

And even more importantly:

👉 how it was decided

The Decision Trace Model (DTM) externalizes the structure of decision-making and provides a framework to:

• Record
• Execute
• Improve

decisions as:

Event → Signal → Decision → Boundary → Human → Log

At this point, a critical shift occurs:

👉 Treating decisions as data

The Decision Trace Ledger enables this by:

• Structuring the flow of decisions
• Preserving temporal and causal relationships
• Storing them in a reproducible form

As a result, it generates:

👉 Traceable Decision Data

In other words, the Ledger:

• Records decision information
• Reconstructs decision processes
• Makes decision histories verifiable

Thus, it becomes:

👉 An infrastructure that turns decision-making itself into an asset

But this raises a fundamental question:

👉 Can this asset be learned?

By combining this with Graph Neural Networks (GNNs), we enter a new phase.

Decision Data accumulated in the Ledger consists of:

• Event
• Signal
• Decision
• Human
• Context

These elements are connected through:

• Causality
• Temporal order
• Dependencies

Forming a:

👉 Graph structure

In other words:

👉 The Ledger is inherently graph data

And GNNs are:

👉 Models that learn patterns and relationships within graph structures

This enables:

👉 Decision-making itself to become a learning target

And this is not just analysis.

👉 It is a mechanism to further increase the value of assetized Decision Data.

■ What GNN Enables

By applying GNNs, decision-making—previously only recorded—becomes:

👉 Learnable, reusable, and continuously evolving

Below are representative use cases:

1. Discovery of Decision Patterns

By clustering similar decision flows, we can extract:

• Successful patterns
• Failure patterns

as structured knowledge.

This makes it possible to visualize what was previously tacit:

👉 “decision habits” and “field expertise”

2. Learning Good vs. Bad Decisions

By linking decisions to KPIs (sales, satisfaction, incident rates, etc.), we can learn:

• Which decision structures lead to success
• Which decisions cause problems

The evaluation target shifts from:

👉 Results → Decision structures

3. Next Decision Recommendation

From the current state (Event / Signal), we can:

👉 Predict the next optimal decision

This is not just generative AI,

👉 but practical decision support grounded in real operations

4. Decision Anomaly Detection

By detecting deviations from normal decision flows, we can identify:

• Fraud
• Early signs of incidents
• Operational deviations

5. Discovery of Causal Relationships

By learning relationships between decisions and outcomes:

• Which decisions contribute to success
• Which decisions introduce risk

can be inferred.

6. Multi-Agent Optimization

DTM involves multiple agents:

• Signal Agent
• Decision Agent
• Boundary Agent
• Human Agent

GNN enables:

👉 Structural analysis of each agent’s influence

7. Decision Importance Analysis

Using graph centrality:

• Identify critical decisions affecting the entire system
• Detect bottlenecks

8. Counterfactual Simulation

With GNN:

👉 Counterfactual analysis becomes possible

Examples:

• What if the order quantity had been different?
• What if escalation had not occurred?

This strongly integrates with Decision Trace Studio.

9. Decision Knowledge Graph

Decisions can be accumulated as reusable knowledge:

• Under these conditions → this decision
• This pattern is risky

👉 Moving from search to decision support

10. Automatic DSL Generation

From learned patterns:

• Conditions
• Priorities

can be extracted and generated as:

👉 Decision DSL

■ The Fundamental Shift

What matters here is not that AI becomes smarter.

What changes is:

👉 What is being learned

Traditional AI:

Data → Prediction

DTM × Ledger × GNN:

Decision Structure → Learning → Improvement

In other words:

👉 Not what was output,
👉 but how decisions are made is learned

■ Decision OS Loop

This structure forms a continuous loop:

Decision Design (Studio)
↓
Execution (Engine)
↓
Recording (Ledger)
↓
Learning (GNN)
↓
Improvement (Studio)

This becomes:

👉 A system where decision-making continuously evolves (Decision OS)

■ Beyond: Decision Embedding

With GNN:

• Decision vectorization
• Similar decision retrieval
• Cross-domain transfer

become possible.

This is:

👉 The first step toward a “Foundation Model for Decisions”

---

■ OSS Implementation: decision-trace-gnn-core

This concept is not just theoretical.

It is being implemented as an open-source project:

👉 Decision Trace GNN Core（decision-trace-gnn-core）

This library:

• Converts Decision Trace Ledger data into graphs
• Learns decision structures using GNNs
• Outputs results usable in real-world systems

Currently supported capabilities include:

• Next Decision Prediction
• Decision Anomaly Detection
• Decision Pattern Clustering
• Decision Embedding

In other words:

👉 It is an OSS that enables not just handling decisions, but learning them

Furthermore, this project evolves as part of a larger ecosystem:

• Decision Trace Ledger
  
• multi-agent-orchestrator-core
  
• Decision Trace Studio

Together forming:

👉 A complete loop of design, execution, recording, learning, and improvement

Importantly:

👉 This is not a finished system, but an evolving foundation

Ongoing improvements include:

• Advanced GNN models (GAT / Temporal GNN)
• Enhanced counterfactual analysis
• Integration with DSL generation
• Multi-agent optimization

Thus:

👉 It is evolving into a system that continuously increases the value of Decision Data as an asset through learning

■ Conclusion

Until now, enterprises have focused on:

• Accumulating data
• Organizing knowledge
• Enabling search

But what truly matters is:

👉 What decisions were made

By combining Decision Trace Ledger and GNN:

Decision-making becomes:

• Recorded
• Analyzed
• Learned
• Improved

And most importantly:

👉 This is not about making AI smarter

It is about:

👉 Changing what is learned

Traditional AI:

Data → Prediction

DTM × Ledger × GNN:

Decision Structure → Learning → Improvement

This means:

👉 Not what was produced,
👉 but how decisions are made becomes the learning target

This loop continues:

Decision Design (Studio)
↓
Execution (Engine)
↓
Recording (Ledger)
↓
Learning (GNN)
↓
Improvement (Studio)

This is:

👉 A continuously evolving decision system (Decision OS)

And beyond that:

• Decision vectorization
• Similar decision retrieval
• Cross-domain transfer

These open the path toward:

👉 A Foundation Model for Decisions

Ultimately:

👉 AI is not about prediction
👉 It is about evolving decision-making systems