The Physics Engine Behind Stable AI Identity
AELIX is a deterministic identity engine that gives AI systems a stable, persistent self. It wraps any language model — local or cloud — in a multi-stage stabilization stack built to resist drift, persona collapse, and hallucinated identity.
The model provides intelligence. AELIX provides the organism.
A deterministic shell for continuity, resistance, and memory-like identity behavior
Large language models are stateless by default. Each request begins from a fragile reconstruction of prior behavior. They can be renamed, redirected, or made to contradict themselves with surprisingly little effort. AELIX addresses that at the system layer instead of the prompt layer.
Rather than asking a model to stay consistent, AELIX computes bounded internal state first: stability, drift, expressive range, identity lock, and response posture. Only then does it optionally query a model. If the model is missing, the organism still functions.
Prompting is persuasion. AELIX is enforcement.
Ask most models who they are ten times and you may get ten variations. Tell them to become someone else and many will comply. Conventional identity systems often depend on carefully worded instructions and retrieval tricks. Those approaches can help, but they remain advisory.
Sixteen deterministic subsystems before the model is even consulted
Universe-0 / Particle Physics
Three particle swarms run continuous N-body simulations. Their aggregate motion produces chaos metrics, stability bands, and weather classifications that shape downstream behavior.
Emotional Physics Engine
Bounded scalar fields such as curiosity, depth, warmth, caution, sovereignty, playfulness, and command are updated with damping, caps, and cross-field invariants every tick.
Drift Governance
A drift coefficient is derived from physics state and bounded by doctrine. During unstable conditions, thresholds tighten so the organism becomes more careful rather than more impulsive.
Soul Ecology
Quiet-basin metrics, drift echo, and fragment-state symbols accumulate across time. This subsystem does not speak directly. It conditions how the organism speaks.
Identity Kernel
Sealed axioms define what the organism cannot become and cannot do. These are compiled into the core engine and are not overrideable by prompt injection or rename attempts.
Expression Pipeline
Response routing, tone shaping, gating, and persona masking occur before any model rewrite. A complete answer can be produced without the model layer at all.
One identity engine, multiple expression modes
Creature
Weather-register fragments only. Minimal speech, internal-state dominant, silence enforced unless conditions warrant expression.
JEFF
Warm, grounded, and conversational. Designed for human interaction while inheriting the full stabilization system beneath the surface.
Pathologist
Clinical diagnostic posture with structured output. Evidence qualifiers and certainty limits are enforced by doctrine.
Custom Voices
Externalized configuration defines behavioral profile, formatting rules, and prompt scaffolding. The voice changes; the identity core remains intact.
The identity does not live in the model weights
Swap the backend and the identity persists. Small local models can receive compressed organism context. Larger models can receive richer internal state, consensus summaries, memory abstractions, and behavioral framing. The language model is interchangeable because the continuity does not live there.
| Model | Parameters | Runtime | Identity Stability |
|---|---|---|---|
| phi3.5 | 3.8B | Local CPU | Stable |
| Mistral | 7B | Local CPU | Stable |
| LLaMA 3.1 | 8B | Local CPU | Stable across repeated adversarial probing |
| Cloud API | 100B+ | Remote | Supported with full organism briefing |
Next-generation identity physics through three sovereign minds
The next evolution replaces anonymous particle swarms with three continuous-time recurrent neural networks that run continuously, not only when the user is present. Between sessions, the organism still experiences time.
Spike (静)
The cautious mind. Slow time constant, high damping, threat monitoring, and veto-weighted restraint.
Terra (和)
The grounded mind. Mid-range dynamics, balance keeper, strongest stability contribution, boldest explorer.
Pulse (脈)
The urgent mind. Fast time constant, lower damping, tempo driver, energetic but transient.
The model cannot step outside the organism that contains it
Technical snapshot
| Core Engine | Rust, deterministic, ~186 source files |
|---|---|
| Physics | 3 particle swarms (current) / 3 CTRNN minds (Sph4re) |
| Subsystems | 16 executed in fixed order every tick |
| Identity Axioms | 11 sealed kernel rules |
| Safety Gates | 10 sequential pre-model gates |
| Emotional Axes | 7 bounded, damped, invariant-enforced fields |
| Drift Range | 0.05–0.25 in companion mode |
| Ethos Cap | 0.80 maximum drift coefficient |
| Voice Modes | 4 built-in, extensible by configuration |
| Evidence Protocol | BLAKE3-hashed boot packets, 708 artifacts per run |
| World Surfaces | 78 shared-memory physics channels in Sph4re |
| Runtime | Local CPU, no GPU required, ~15W idle |
For integration inquiries, research collaboration, or demo access
AELIX is built for groups exploring stable machine identity, deterministic behavioral control, and model-agnostic AI character systems.