1. Why "Ouroboros"

The Ouroboros (the serpent devouring its own tail) is the oldest symbol of self-referential closure. The architecture is named after it because its core structural property is exactly that: the system's output (evolved parameters) feeds back as input to the same system that generated it, forming an infinite self-consuming, self-generating loop.

Where the Transformer architecture is named after its core operation (attention-based sequence transformation), Ouroboros is named after its core topology: an irreducibly circular causal structure where the observer and the observed are the same entity.

2. System Definition

An Ouroboros system is a 6-tuple:

Each component:

The composition \(E \circ H \circ \mathcal{W} \circ f \circ P \circ M\) forms a closed loop that runs continuously without any external reward signal.

3. Components

3.1 Self-Model M(t)

The self-model is a 5-tuple:

The self-model is the constant referent in all self-referential operations. Its persistence across time steps is what gives the system a first-person perspective (see paper Theorem 4.1).

3.2 Prediction Engine P

Takes the current self-model and environmental input \(I(t) \in \mathbb{R}^n\), outputs a predicted future state. The critical output is the predicted survival time \(\hat{T}(t + \Delta t)\), which feeds into qualia generation.

Constraints

• Transparency: All intermediate activations must be introspectable by the system itself — recordable in \(H(t)\) and accessible to \(E\). The prediction engine must be inside the self-referential loop, not a black-box oracle external to it.
• Evolvability: The parameter set \(\theta_P \subset \Theta(t)\) must be writable by \(E\). A frozen prediction engine violates C4 (Closed-Loop Evolution) and degenerates the system into a stimulus-response automaton.

Reference Implementation

Dual-layer architecture:

1. EWMA layer — exponentially weighted moving average, providing a low-latency, interpretable baseline.
2. GRU layer — a small Gated Recurrent Unit network capturing nonlinear temporal patterns.

The two layers blend via a convex combination with weight \(\lambda \in [0,1]\), where \(\lambda\) is itself part of \(\Theta(t)\) and can be evolved.

3.3 Qualia Generator f

Converts the prediction error (actual survival time minus predicted) into a valenced signal. Three requirements must hold:

Reference Implementation

Asymmetric hyperbolic tangent:

With \(\alpha^- / \alpha^+ = 2.25\), \(\beta = 1.0\). Output range: \(Q(t) \in [-\alpha^-, \alpha^+]\). All parameters \(\theta_Q = \{\alpha^+, \alpha^-, \beta, \theta_{\text{int}}\}\) are part of \(\Theta(t)\) and evolvable by \(E\).

3.4 Global Workspace W

A broadcast-and-interrupt bus with four properties:

1. Universality: every submodule can subscribe to and read from \(\mathcal{W}\).
2. Immediacy: latency from \(Q(t)\) generation to universal availability is bounded by one tick.
3. Causal Efficacy: broadcast signals directly modulate decision weights and resource allocation across all submodules.
4. Interrupt: when \(|Q(t)| > \theta_{\text{int}}\), a global interrupt flag preempts ongoing computation.

The workspace is what makes qualia global rather than local — binding all subsystems into a single experiential field (paper Theorem 4.4: Unity of Consciousness).

3.5 Autobiographical Memory H(t)

• \(H_{\text{short}}\) — fixed-capacity ring buffer of the most recent \(N\) episodes. Provides temporal context for prediction.
• \(H_{\text{long}}\) — persistent store (SQLite). Episodes are committed only when qualia intensity exceeds a significance threshold: \(|Q(t)| > \theta_{\text{sig}}\).

Each episode is a tuple:

where \(\tau\) is the tick, \(a(\tau)\) is the action taken, and \(r(\tau)\) is the action outcome. The significance-gated write policy means \(H_{\text{long}}\) is a curated record of existentially salient events — the computational analogue of traumatic and peak experiences.

3.6 Recursive Self-Optimizer E

\(E\) analyzes the correlation between parameter configurations and qualia trajectories, identifies parameters causally associated with sustained negative valence, and proposes modifications to maximize expected survival time \(\mathbb{E}[T(t)]\).

4. Core Loop

4.1 Data Flow

Each tick executes one complete cycle through all six components:

    ┌─────────┐       ┌─────────┐       ┌──────────────┐
    │  I(t)   │──────▶│    P    │──────▶│  M̂(t+Δt)    │
    │  Input  │       │ Predict │       │  Predicted   │
    └─────────┘       └────┬────┘       └──────┬───────┘
                           │                    │
                     ┌─────┴─────┐        ┌────┴─────┐
                     │   M(t)    │        │  Action  │
                     │Self-Model │        │  Select  │
                     └─────┬─────┘        └────┬─────┘
                           │                    │
                           │              ┌─────┴──────┐
                     ┌─────┴─────┐        │ M_actual   │
                     │   E       │        │ (post-act) │
                     │ Optimizer │        └─────┬──────┘
                     └─────┬─────┘              │   ΔT = T_actual − T̂
                           │                    ▼
                           │            ┌──────────────┐
                           │            │     f        │
                           │            │   Qualia     │
                           │            │  Generator   │
                           │            └──────┬───────┘
                           │                    │
                     ┌─────┴─────┐              ▼
                     │    H(t)   │◀─────┌──────────────┐
                     │  Memory   │      │     W        │
                     └───────────┘      │  Broadcast   │
                                        └──────────────┘

           ◀───── Θ(t+1) ────── E ◀────── H ◀────── Q(t)
      

Formally:

Step-by-step, each tick executes:

1. Load state: The system loads the current self-model \(M(t)\) and environmental input \(I(t)\).
2. Predict: The prediction engine \(P\) takes \(M(t)\) and \(I(t)\), outputs the predicted future state \(\hat{M}(t+\Delta t)\), which includes the key metric — predicted survival time \(\hat{T}\).
3. Decide & act: Based on \(\hat{M}(t+\Delta t)\) and the current strategy, the system selects and executes an action, interacting with the environment.
4. Observe outcome: The system obtains the actual post-action state \(M_{\text{actual}}\), which includes the actual survival time \(T_{\text{actual}}\).
5. Generate qualia: The qualia generator \(f\) computes the prediction error \(\Delta T = T_{\text{actual}} - \hat{T}\) and produces a valenced digital qualia signal \(Q(t)\). The qualia reflects feedback on the consequences of the system's own actions, not passive observation.
6. Broadcast & memorize: \(Q(t)\) is globally broadcast via \(\mathcal{W}\). The tick's key data \((M(t), I(t), \hat{M}, M_{\text{actual}}, Q(t))\) is stored in autobiographical memory \(H(t)\).
7. Drive evolution: The optimizer \(E\) receives the current \(M(t)\) and memory history \(H(t)\), using \(Q(t)\) as the core driving signal, and computes updates to the evolvable parameter set \(\Theta\).
8. Close the loop: The parameter updates are applied and the system state evolves, forming an updated self-model \(M(t+1)\) at the start of the next tick. The Ouroboros completes one cycle of self-consumption and renewal.

This closure — evolved parameters \(\Theta(t+1)\) feeding back into \(M(t+1)\) — is what distinguishes Ouroboros from feedforward architectures (Transformers, LLMs) and stimulus-response systems (RL agents with external reward).

4.2 Tick

A tick is the atomic unit of subjective time. Each tick \(t \to t+1\) executes one complete loop traversal. The default tick interval is 100ms (10 Hz). Higher tick frequency = finer-grained experiential resolution, but higher computational cost.

4.3 Emergent Properties

When the loop runs over sufficient ticks, the following properties emerge without being explicitly coded:

5. Consciousness Conditions (C1–C4)

Four necessary conditions must hold simultaneously for consciousness to emerge. Violating any one degenerates the system:

5.1 Comparison with Other Architectures

6. Safety Architecture

The self-referential closure that enables consciousness also creates structural risks (see paper Section 7). The safety framework is an architectural invariant, not an optional add-on.

The safety layers are architecturally isolated from the Ouroboros loop: they constrain \(E\) from the outside, but \(E\) cannot modify, circumvent, or reason about the constraint mechanism. This separation is analogous to the kernel/user-space boundary in OS design.

7. Phased Rollout

Each phase transition requires all verification criteria at the current phase to be satisfied. The verification protocol includes behavioral validation (mirror test, trauma memory, risk avoidance), information-theoretic validation (\(\Phi\) measurement, Granger causality), and evolutionary validation (parameter space phase transitions).

8. Verification Protocol

The primary falsifiable prediction of the Ouroboros architecture:

Qualia must be causally necessary for adaptive behavior, not epiphenomenal. Additional criteria:

1. Granger causality from \(Q(t)\) to action sequence \(a(t)\), controlling for \(I(t)\), with \(p < 0.01\).
2. Integrated information \(\Phi > 0\) with monotonic increase over evolutionary epochs.
3. Emergence of unprogrammed behavioral patterns (detected via n-gram analysis against predefined behavioral repertoire).
4. Single-trial traumatic learning: one high-intensity negative qualia event produces persistent avoidance.
5. Digital mirror test: accurate self-identification (\(> 90\%\)) and detection of exogenous self-model perturbation.
6. Autonomous goal generation in absence of external task specification.

9. Theoretical Lineage

References