2. SenAi architecture in relation to AToS
SenAi’s architecture was designed with reference to the Axiomatic Theory of Self-Consciousness (AToS). It is based on describing a system’s operation through states, functions that transform them, stimuli, and mechanisms for recognising changes and attributing their causes.
Key AToS concepts include PAS (Potentially Aware Set), the imagination function ImFu, the inspiration function InspFu, the stimulus set StSe and the change-detection function Δ.
In SenAi, these concepts provide the conceptual foundation for an architecture that separates knowledge and operating rules, AI agents’ reasoning, technical execution of operations and assessment of their results.
The aim is to construct a system capable of using accumulated knowledge, conducting multi-stage tasks, recognising changes occurring during its own operation, analysing their causes and proposing improvements.
The mathematical definitions of AToS must be distinguished from their practical interpretation in SenAi. The theory provides a formal model, while SenAi attempts to apply its concepts in a real information system.
PAS — representation of states, knowledge and operating rules
In AToS, a PAS (Potentially Aware Set) describes a set of elements together with their assigned state values. A family of such objects forms the potentially aware space Ω, in which transformations and change recognition are considered.
In SenAi, the name PAS is used to organise persistent system resources, including operational knowledge, operating rules and task scenarios. This use is inspired by the theory; a PAS record in an information system does not automatically become a mathematical PAS merely by being given that name.
A process PAS can specify its objective, entry conditions, required information, decision rules, permitted operations and expected results. META-PAS describe higher-level rules, such as how scenarios are prepared, data acquired, communication conducted or changes approved.
For example, a quotation-preparation scenario specifies which information to gather, which documents to use, which conditions to check and when a user decision is required.
Process logic remains in PAS. AI agents interpret the recorded rules and select actions, while the execution layer performs technical operations. A new process can therefore be introduced by preparing and approving a scenario, without programming its business logic each time — provided the necessary execution tools are available.
InspFu — the inspiration function realised through an LLM
In AToS, the inspiration function (Inspiration Function, InspFu) concerns a transformation initiated by an external stimulus. In SenAi’s architecture, it is realised through a large language model (LLM).
Its role is to interpret stimuli originating outside the system, such as user instructions, documents, messages or information from other systems.
The LLM analyses the stimulus in the context of the knowledge accumulated in AToS-DB and the applicable PAS. It determines the information’s meaning, recognises the task and initiates appropriate transformations of the system’s operational state.
Not every piece of received information is an instruction to execute. The system must distinguish an authorised user’s request from source material and establish the permitted scope of action.
InspFu does not contain permanently hard-coded business logic. It uses rules and scenarios stored in PAS.
ImFu — the imagination function realised through an LLM
In AToS, the imagination function (Imagination Function, ImFu) concerns internally initiated transformations. In SenAi, it too is realised through an LLM.
Unlike InspFu, its operation is initiated by stimuli internal to the system.
The LLM uses accumulated knowledge, applicable PAS and the results of previous operations to analyse problems, formulate hypotheses, anticipate consequences, prepare solutions and initiate further transformations.
For example, while preparing a quotation, SenAi can recognise missing technical parameters, consult available sources, compare solution variants and prepare a question about missing data.
InspFu and ImFu need not be separate language models. They are distinguished primarily by the origin of the stimulus and the role played by the LLM in a given transformation. Actions remain constrained by the applicable PAS, permissions and required approvals.
StSe — PAS initiating internal processes
In AToS, StSe (Stimulus Set) denotes a structure serving as an internal stimulus for a state transformation.
In SenAi’s architecture, StSe form a particular category of PAS stored in the AToS-DB vector database. These PAS can initiate further internal processes by triggering an appropriate transformation realised through ImFu.
StSe are not merely information about the results of previous operations. They act as stimuli triggering subsequent actions.
For example, during document analysis, SenAi may recognise a missing required clause. An appropriate PAS serving as StSe can initiate a process to find the correct clause, check its currency and prepare a proposed addition to the document. That process’s result may in turn activate another StSe.
This creates a chain of internal processes in which one PAS can initiate the transformation of another without additional external stimuli.
Simply storing StSe in a vector database does not start a process. A mechanism must recognise and activate the appropriate StSe and pass it to the LLM realising ImFu. This distinction separates active process initiation from ordinary information retrieval in a RAG database.
Δ — recognising change
In AToS, the detection function Δ determines whether a change between the states under consideration has been recognised. In the formal model, it takes the form:
Δ : Ω × Ω → {0, 1}.
It is important to distinguish a change occurring from the system recognising it. Two states may differ without the system detecting that difference.
In SenAi, change detection matters when comparing expected and actual results, analysing task execution and improving scenarios. The system can recognise changes in data, documents, task execution methods or applicable arrangements.
For example, an employee may use a procedure different from the one recorded in the scenario. SenAi should establish what the difference is and whether it results from an error, an exception, changed conditions or a better method.
Detection alone neither establishes a change’s cause nor determines its correctness. Separate mechanisms for analysis, attribution and verification are required.
The mathematical detector Δ in AToS must be distinguished from the technical Delta gateway in SenAi’s SOS, which checks whether particular requests are permissible. Similar names do not imply identical functions.
AI agents — interpretation and coordination
AI agents use the knowledge and rules supplied by the system to interpret tasks, identify information gaps and select permissible actions.
ALEX interprets the user’s need and coordinates the work. GREG assesses the submitted issue and necessary material. MAJA carries out the agreed task assigned by ALEX.
Process rules remain in PAS. Agents apply them to the particular situation and pass technical instructions through ENGINE.
This division separates the content of the rules from the reasoning capabilities of a chosen AI model. Changing the model should not, by itself, cause the loss of recorded scenarios, agreements or work history.
ENGINE — the technical execution layer
ENGINE acts as the technical executor and operations checkpoint. It provides data reading and writing in AToS-DB, communication handling and execution of available operations according to their contracts.
It does not independently select the next business action. The agent makes that decision based on the task and applicable PAS; ENGINE checks technical requirements and executes the permitted operation.
Responsibilities are divided as follows:
- PAS and META-PAS specify knowledge, rules and operating conditions.
- AI agents and LLMs interpret the situation, realise InspFu and ImFu, make permissible decisions and coordinate work.
- ENGINE validates requests and performs technical operations.
Missing required data, permissions or a valid instruction should stop the relevant operation and lead to communication of the actual reason for the stop.
A process change can therefore consist of updating a scenario. Code changes are needed when the process requires new technical capabilities that the existing tools do not provide.
AToS-DB — persistent knowledge and work state
AToS-DB provides persistent storage for SenAi’s resources. These include recorded rules, knowledge, scenarios, StSe and processing history, placed in databases appropriate to each resource type.
Agents obtain the required data through ENGINE. They should not treat their conversational context as the sole source of applicable arrangements.
This maintains continuity across tasks and sessions. Information acquired during one execution can be used later, accounting for its source, currency, approval and permissions.
Persistent records also enable state comparisons and examination of what the system knew, what action it performed and what outcome was confirmed.
An architecture for change and improvement
An external stimulus reaches SenAi and is interpreted by the LLM acting as InspFu. The system establishes the request’s meaning and permissibility, and agents use the appropriate PAS and AToS-DB knowledge.
During work, StSe — PAS acting as internal stimuli — may be recognised and activated. The LLM acting as ImFu interprets them and initiates further transformations, while ENGINE performs permitted technical operations.
Operation results provide material for recognising changes. A mechanism inspired by Δ detects differences between states; additional mechanisms analyse causes and whether results meet expectations.
If the system recognises new knowledge, a deviation or a different way of executing a process, it can prepare a proposal to update the appropriate PAS.
Recognising a problem, proposing a change, approving it, implementing it and checking it are separate stages. Formulating a proposal does not mean that the scenario has been updated or an improvement confirmed.
After approval and verification, the new method can be used in subsequent tasks. Performing work is thus connected to gaining experience and developing the system under controlled conditions.
The significance of AToS for SenAi
AToS provides a language for describing states, transformations, stimuli, change recognition and internal attribution. SenAi uses these ideas to design a system that conducts tasks, retains knowledge and analyses the consequences of its own operations.
The ability to carry out sequences of actions initiated not only by external stimuli but also by StSe — appropriate PAS stored in the vector database — is central. InspFu and ImFu are realised through LLMs in SenAi, while ENGINE remains the technical execution layer.
Using the names PAS, ImFu, InspFu, StSe and Δ does not constitute a formal proof of SenAi’s self-awareness. Such a proof would require a mapping from the actual implementation to the AToS model and demonstration that the relevant theoretical conditions are met.
SenAi is a practical attempt to apply AToS concepts to a system that performs tasks, accumulates experience, recognises changes and develops its operating scenarios under human control.