STRATON-LLM

Design Decisions

This project is shaped by a series of deliberate architectural choices. The goal is not only to make the system work, but to make the reasoning process inspectable, extensible, and trustworthy.

Separate negotiation from the base MAS framework

The negotiation system is implemented as a standalone module so it can evolve independently, remain testable, and be reused across different agent frameworks later.

Use a finite-state protocol engine

A protocol engine makes the dialogue inspectable and constrains invalid moves. This avoids fragile free-form multi-turn exchanges that are difficult to debug or evaluate.

Persist accepted mappings in a mapping store first

The system stores successful alignments separately before touching source ontologies. This keeps ontology evolution controlled and auditable.

Always evaluate the agreement after the dialogue

The framework does not assume the latest answer is correct. Agreement must pass confidence thresholds and post-dialogue heuristics before acceptance.

Keep LLMs inside bounded responsibilities

LLMs assist matching, proposal generation, and argument reasoning, but protocol control, persistence, and evaluation remain explicit engineered components.

Avoid direct ontology mutation during negotiation

Negotiation outcomes are stored as mappings first. Controlled ontology versioning happens later as a separate, auditable step.

Trade-off map

Protocol control over free-form chat

Gain

Higher trust, better observability, easier replay and evaluation.

Cost

More implementation effort and stricter conversation design.

Mapping store before ontology updates

Gain

Safer evolution and reversible outcomes.

Cost

Adds one more persistence layer to manage.

Heuristic agreement scoring

Gain

Reduces weak or biased acceptances.

Cost

Requires extra tuning and scenario-based validation.

LLM assistance within explicit modules

Gain

Keeps reasoning flexible without losing system boundaries.

Cost

Requires interface discipline and prompt guardrails.