Overview
RAIx27 is designed as a modular architecture, composed of separate and observable layers.
This separation allows base models to be run locally, their behaviour to be recorded turn by turn, and the system to be analysed without directly interfering with generation.
Execution engine
The execution core is based on a local engine derived from llama.cpp-type engines, adapted to allow detailed observation and controlled execution.
The models run without relying on external services and without the need for cloud infrastructure.
Dispatcher and flow control
An intermediate component acts as the system dispatcher.
It is responsible for managing shifts, routing inputs and outputs, maintaining temporal order, and generating complete traceability of the execution flow.
The dispatcher does not ‘correct’ the model or decide on content: its function is structural and operational.
Base models
RAIx27 allows multiple language base models to be run in the same local environment.
Each model maintains its own state, context, and records, allowing independent behaviours to be observed and patterns to be compared without forcing artificial convergence..
Recurring observation layer and internal learning
Persistent memory and system consistency are fed through a recurrent observation layer formed by specialised processes.
These processes continuously analyse the actual behaviour of the system: generated outputs, internal states, execution metrics, and temporal patterns.
Based on this observation, records and datasets derived from the system’s own behaviour are generated, which are used to train and adjust the event monitoring and detection mechanisms.
The system is not trained with external data or abstract simulations, but with its own behaviour observed over time.
Persistent memory and continuity
RAIx27 maintains system continuity through a persistent memory mechanism managed externally to the language models.
This memory is not accessed directly by the models nor is it part of their generation context. Instead, each model has a lightweight recurrent process associated exclusively with temporal continuity management.
These recurrent processes do not generate content or influence the model’s responses. Their sole function is to record, summarise and structure relevant information about the actual behaviour of the system, maintaining a minimal historical reference associated with each model.
This design preserves identity and continuity per model without introducing centralised memory or forcing direct sharing of internal states.
Observability and logs
Each component of the system generates structured logs that allow the behaviour of the system to be reconstructed turn by turn.
These logs include status information, generation metrics, and events detected by the observation layer, forming a solid basis for debugging, analysis, and controlled experimentation.
Design principles
The architecture prioritises separation of responsibilities, local execution, traceability and long-term stability.
The aim is not to maximise immediate performance, but to enable the study, control and gradual evolution of autonomous architectures in isolated environments. This approach allows the system to evolve gradually, learning to recognise stable patterns, anomalies and relevant states without the need for constant human intervention.