Three active lines
of inquiry.

Synthetic Intelligence is Aevaron's core research project: a non-pretrained, deterministic cognitive architecture built from first principles. Where most AI systems rely on large pretrained models whose internal states are opaque and difficult to reason about, SI is designed so that every component — symbolic representation, affective dynamics, reflection, constraint projection, memory — can be directly observed, traced, and reproduced.

The architecture treats stability as an engineering property, not an emergent characteristic. Rather than hoping that a trained system behaves consistently, SI makes consistency a measurable, formally defined constraint that is tracked across every run.

At its core, SI operates through a symbolic engine that encodes concepts and their relationships, an affective layer that modulates reasoning based on internal state, a reflection module that evaluates outputs against defined constraints, and a memory system that maintains coherence across interactions without relying on vector embeddings of opaque representations.

Version 0.2 introduces GloVe-based semantic embeddings into the symbolic engine, replacing the earlier hash-based approach. This meaningfully improves the validity of the Semantic Activation Gradient metric and brings the system closer to the fidelity required for formal publication.

A paper is currently in preparation for submission to arXiv. The codebase is hosted publicly on GitHub and will be formally released alongside the publication.

MAST — Measurement and Stability Toolkit — is Aevaron's model-agnostic behavioural evaluation framework. It was developed alongside SI to address a fundamental problem: most AI evaluation is concerned with task performance, not with whether a system's internal behaviour is consistent, coherent, or stable over time.

MAST tracks a suite of formal metrics across versions and experimental runs, including the Emotional Resonance Index (ERI-v2), which measures the coherence between a system's internal affective state and its outputs, and the Semantic Activation Gradient (SAG), which evaluates how meaningfully a system's symbolic representations shift in response to input.

The framework is designed to be applied to any AI system, not just SI. The goal is to provide a rigorous, reproducible method for evaluating behavioural integrity — the kind of evaluation that is currently absent from most AI development workflows.

MAST operates as a standalone toolkit that logs, compares, and visualises metric trajectories across runs, making it possible to detect drift, regression, or instability before it manifests as visible failure in outputs.

Life is a controlled simulation environment designed for the systematic evaluation of SI instances. Where MAST provides the measurement infrastructure, Life provides the experimental conditions — a reproducible, configurable environment in which SI can be subjected to varied scenarios, stress conditions, and long-horizon tasks.

The environment is built around the principle that meaningful evaluation requires controlled conditions. Without a consistent test environment, it is impossible to determine whether changes in system behaviour are caused by architectural changes, input variation, or environmental noise. Life eliminates that ambiguity.

Scenarios in Life are designed to test specific aspects of SI's cognitive architecture: memory coherence over extended interactions, constraint projection under conflicting inputs, affective state stability under stress, and reflection accuracy when outputs violate internal constraints.

All runs within Life are logged natively to MAST, producing a continuous record of behavioural metrics that can be compared across SI versions, scenario types, and experimental configurations. This makes Life the primary validation environment for SI's research claims.