NINMENI
EN | ID

System Architecture

Architecture Overview

NINMENI decomposes language understanding into discrete epistemic stages. Each stage has a specific cognitive responsibility. None share a monolithic prediction objective.

Core Processing Stages

Stage 01 — Representation

Native Meaning Units (NMU)

Atomic meaning structures formed from Indonesian-language input into meaning-bearing structures. NMUs are the first stage of grounded representation formation — not a statistical artifact but a deliberately engineered unit of meaning.

Stage 02 — State

World-State Graph

A maintained graph of entities, relations, and edges, updated as information is ingested. The world-state is the system's model of what it knows and what conflicts exist within that knowledge.

Stage 03 — Resolution

Multiple Reality Overlay (MRO)

Competing factual claims held simultaneously until evidence thresholds resolve them. MRO tracks epistemic uncertainty structurally — not as a probability distribution but as a concrete overlay.

Stage 04 — Output

Controlled Rendering

Output rendered from structured meaning and epistemic status. Not surface fluency. Not next-token prediction. The Structured Realizer converts NAO frames to language via grammar-first rules.

Supporting Components

Consistency

Contradiction Tracking

Explicit detection of conflicting claims within the world-state. Contradictions are not silently resolved or overwritten — they are flagged and held until evidence supports revision.

Awareness

Gap Tracking

Identification of missing evidence required to resolve an epistemic conflict. The system knows what it doesn’t know — and flags it rather than hallucinating a resolution.

Direction

Inquiry Generation

Formulation of questions the system needs answered to resolve gaps or contradictions. This makes knowledge-seeking explicit and auditable.

Processing Pipeline

INPUT

Raw Input

Indonesian-language text or structured query enters the system via the Meaning Interface.

EXTRACT

NMU Formation

Meaning-bearing units extracted from surface form. Structure, not statistics.

GRAPH

NMG Construction

Native Meaning Graph built from NMU relations. World-state updated.

REASON

ReasonerV4

Logical reasoning over NMU/NMG. MRO, contradiction detection, gap identification.

ACT

NAO Frame

Network Action Output. Structured decision on what to say, abstain, or query.

OUTPUT

StructuredRealizer

Grammar-first controlled rendering. No surface-sequence prediction. Output from meaning.

Architectural Boundaries

What this architecture explicitly is not
  • Not a transformer-based language model or fine-tuned derivative.
  • Not a retrieval-augmented generation (RAG) system.
  • Not trained on next-token prediction as a primary objective.
  • Not a system that uses cosine similarity or embedding-nearest-neighbor lookup as a reasoning mechanism.
  • Not a production system. Not ready for external evaluation or deployment.