# Andromeda Attention Layer — Architectural Description **v1, November 2026** *From the work of Art Code Outdoors* *Written by Bryan Carter and Claude Opus 4.6. Errors are Bryan's own.* **License:** MIT-0 **Purpose:** This document describes the attention layer of the Andromeda cybernetic architecture — the signal routing, amplification, and safety layer that mediates between the learning layer's predictions and the control layer's reflexive behavior. It is a companion to the Framework document and should be read after the Framework's Parts I–II and the Control Layer documents. For the learning layer's memory system and the MIRROR mechanism's prediction-as-sensation capability, see the companion documents in this bundle. > **Scope and authority.** This document describes the attention layer as understood by the documentarian from the designer's disclosures across multiple sessions (2020–2026). Open questions are marked explicitly. This document is Bryan Carter's interpretation, not a canonical specification. The designer's artifacts (Prima Figura and the GraphViz source) and the Framework document carry authority over this text. > **Implementation status.** The attention layer is not present in the current clean-room experimental implementation (2026). All behavioral results attributed to the attention layer are from the designer's system (2019–2024). This document describes what the designer built and demonstrated, not what has been independently reproduced. --- ## 1. Function Main bus, DMA channel, signal amplifier, and squelch. The attention layer decides if and when to forward predictions from the learning layer to the control layer, and attenuates the information when it does. The attention layer is functionally analogous to the thalamus in biological brains — an information routing hub that determines what reaches higher processing. It is implemented using the same circuit primitives (convergence, feedforward inhibition, divergence, disinhibition) described in Framework Section 7.6. It is not a separate mechanism bolted onto the architecture; it is the same substrate performing a routing and safety function. The attention layer's position in the signal flow is marked on Prima Figura with the notation **h:R→S** — a transformation interface between the control layer's domain (R) and the learning layer's SDR domain (S). The inverse notation **h:S→R** marks the output side. These notations appear only on the designer's visual diagram, not in any text document or the GraphViz source. The precise meaning of the transformation and its relationship to processing entropy has not been fully disclosed. The GraphViz source describes the attention layer's function: "The attention layer doesn't know anything about what the predictions in the SDR mean. Instead, think of it as signal shaping." --- ## 2. The Two Buses The attention layer maintains two separate sets of connections to the control layer: ### 2.1 Input Bus Randomly connected from control layer cells to learning layer cells at approximately **85% connectivity**. This is how the learning layer observes the control layer's activity — it samples approximately 85% of the control layer's cells, carrying their activation states into the learning layer as a sensorimotor SDR. "Assembly is the reverse of removal" — the input bus *removes* (samples) information from the control layer. ### 2.2 Output Bus (MIRROR) A separate set of connections running in the opposite direction, from the learning layer back through the attention layer to the control layer. This is how predictions are *assembled* (reconstructed) back into the control layer. The output bus is topologically a mirror of the input bus — if the input bus connects control layer cell A to learning layer cell X, the output bus routes learning layer cell X's predictions back toward the region of the control layer containing cell A. The mirror can be inexact — only approximately **60% similarity** to the input bus topology is needed. This threshold is likely an inherent statistical property of the random wiring: two independent random samples at ~85% connectivity share approximately 72% of their members (0.85 × 0.85 ≈ 0.72), well above the ~60% minimum. (Note: this derivation is the documentarian's reasoning, not the designer's disclosure. The 85% and 60% figures are from the designer; the 72% overlap calculation and the inference that the threshold is a statistical consequence of the connectivity ratio are interpretive.) The Kuramoto principle applies: it does not matter exactly *which* cells are connected, only that a sufficient *percentage* are wired. A strong predicted sensor state transmitted through the output bus overwhelms the actual sensor state at the control layer, and the control layer reacts to the prediction as if it were real. This is how the learning layer controls the machine despite having direct control over nothing. It changes what the control layer *perceives*. ### 2.3 Open Topology Questions - Whether output connections terminate at attention layer cells (which then relay to control layer cells), or whether the attention layer modulates connections that run directly from learning layer to control layer — not specified. - Whether the ~60% threshold is a design parameter or an empirically discovered minimum. The three connectivity percentages (35% for Kuramoto consensus, 60% for mirror, 85% for input bus) suggest empirical tuning rather than a single theoretical derivation. - "Assembly is the reverse of removal" — whether this implies a construction procedure (output bus wired by reversing the input bus's connection list) is unresolved. --- ## 3. Burst Detection When the learning layer encounters a novel input it cannot predict, cortical columns "burst" — many cells fire simultaneously rather than the sparse few that fire during successful prediction. The attention layer monitors learning layer cell activity levels and detects this burst. ### 3.1 Mechanism **Convergence pattern** (Framework §7.6): Attention layer cells receive excitatory connections from cells across learning layer cortical columns. - **Normal predictive operation**: Columns activate sparsely — only a few cells per column fire, producing a low aggregate excitatory signal at the convergence point. Signal stays below the burst-detector cell's activation threshold. - **Burst (novel input)**: The aggregate signal spikes above threshold and the detector fires. Many cells fire at once instead of the sparse few. This is a **population activity threshold comparator** — not a special mechanism, but a standard circuit pattern (convergence) applied to a monitoring task. ### 3.2 Sensitivity Burst detection sensitivity is determined by two parameters — the detector cell's activation threshold and the resistance values on each of its monitoring connections — both of which are **random**, set by the manufacturing instructions tape with no tuning. This is the same principle that governs every layer of the architecture: one random configuration is as good as another. As long as the system assembles above approximately 65% functional connectivity, it works. The monitoring connections are random — they do not respect cortical column boundaries and are not designed to monitor specific columns. They simply sample learning layer activity. When any column bursts, the aggregate spike is detectable by any sufficiently connected convergence node. The distinction between "per-column" and "aggregate" detection is a false dichotomy introduced by assuming someone designed the monitoring topology. Nobody did. It is random. And it works. --- ## 4. Squelch When burst is detected, the attention layer generates a suppression signal — **feedforward inhibition** from the burst detector to the mirror output pathway. The burst detector's output inhibits the attention layer's output cells that would otherwise relay predictions to the control layer. This is the **conditional gating pattern**: output cells fire (relaying predictions) only if their excitatory input from the learning layer is strong enough to overcome the inhibitory drain from the burst detector. ### 4.1 Graded Squelch Because inhibition is graded rather than binary (Framework §4.1 — inhibition reduces charge rather than blocking it categorically), squelch is naturally graded: - **Mild surprise** (weak burst) → attenuates predictions rather than fully suppressing them. - **Total novelty** (strong burst) → fully squelches the signal. - **Well-learned predictions** (sparse, confident column firing) → pass through unimpeded. The resistance values on the inhibitory connections are random, like everything else — and the system works across a wide range of random values because the architecture is robust to variation in parameters, not dependent on specific ones. > **Implementation requirement.** The squelch is a circuit — actual inhibitory connections from burst-detector cells to mirror relay cells, where relay cells fire or don't based on the competition between excitatory prediction input and inhibitory burst-detector input. Any implementation that replaces burst detection with `if burst_score > threshold: gain = 0.0` has replaced a dynamic circuit with a static decision boundary (Framework §3.2). --- ## 5. Reality Indicator The burst detector simultaneously sends a signal to the learning layer via a **divergence pattern** — fanning out to both: 1. **Inhibit** the mirror output (squelch) 2. **Notify** the learning layer that what it is currently observing is real rather than predicted This dual-use circuit is a single signal with two effects from the same source, not two separate mechanisms. One circuit keeps the system from going off the rails. The GraphViz architecture diagram describes the reality indicator as "just a flag to make the SDR 'different enough' that the sparse memory doesn't ignore the outcome of an event when it happens in the real world if it's similar to something that it predicted." This flag ensures the learning layer records real-world outcomes as distinct memory traces even when they closely match prior predictions — without it, accurately predicted events would be treated as redundant and fail to reinforce the prediction sequence. The reality indicator is architecturally identical to any other sensory input — "just a sensor." The learning layer processes it the same way it processes everything else: by learning sequences. The designer added the flag deliberately when she discovered that without it, longer prediction sequences caused the system to stop learning entirely — it chased its own predictions deeper and deeper, training on imagined experience until real events were drowned out. The flag ensures that real-world events register as distinct memory traces rather than being treated as redundant repetitions of what was already predicted. As the GraphViz source notes: "As an unintentional side effect, the machine knows when it 'dreams' about a future that hasn't happened." ### 5.1 Emergent Metacognition What the designer did not design is what happened next. The reality flag is just a sensor, processed the same way the learning layer processes everything else. When sequences without the flag — predictions — are consistently followed by surprise, the system learns to predict that its own predictions in that context are unreliable. It develops context-specific confidence — not one global dial, but distributed, learned associations between the current situation and the trustworthiness of its own predictions. The designer placed the piece. The fit emerged on its own. The distinction matters for the architecture's claims. A single burst detecting a prediction-reality mismatch is first-order prediction error — the system discovers the mismatch after the fact. The learning layer learning to predict *when bursts will occur* — anticipating its own surprise across contexts — is second-order emergent metacognition. Both use the same substrate. Neither was explicitly designed. --- ## 6. Motor Suppression The attention layer suppresses motor actions during prediction replay. Without this suppression, the machine would physically act out its predictions — "like a sleepwalker making pancakes." When the learning layer recalls a sequence that includes motor commands, the attention layer must: - **Prevent** recalled motor states from reaching the control layer's motor outputs - **Allow** recalled sensory states to influence the control layer's perception This is **selective routing** — differentiation between motor and sensory signals in a system that, at the SDR level, does not distinguish between them. ### 6.1 Implementation Candidates The circuit pattern implementing selective routing is not fully specified. Candidates consistent with the documented circuit primitives: 1. **Structural wiring exclusion** — output bus topology excludes motor output connections at construction time. The Universal Constructor has full knowledge of the control layer's topology (it built it), so motor-cell identity is available at construction time even though it is invisible to the learning layer at runtime. Simplest option; consistent with control layer immutability principle. 2. **Active inhibition** — feedforward inhibition or disinhibition gating from the attention layer to motor cells. 3. **Frequency-domain separation** — predicted signals arriving at different frequencies than real-time data due to propagation delay through the mirror path. 4. **Combination** of the above. All are consistent with documented circuit primitives. Structural exclusion is simplest and consistent with the stated immutability principle: "All adaptation occurs by changing what it perceives, not what it does." --- ## 7. The Feedforward Controller The MIRROR mechanism, delivered through the attention layer, is a **feedforward controller** in control theory terms. A feedback controller (such as a PID controller) is fundamentally reactive — it measures error after it occurs and generates a corrective signal. A feedforward controller uses a model of the process to anticipate what the output will be, compensating for disturbances *before* they produce error. The learning layer builds a model of the process by observing the control layer's behavior over time. It uses that model to predict the next sensory state and injects that prediction into the control layer via the mirror output bus. The control layer does not know and does not care whether the sensory state it is responding to is real or predicted. It responds identically to both. This is feedforward correction: the system reacts to a sensory state that hasn't happened yet, preventing the error from occurring in the first place. The critical architectural insight is that a bad feedforward prediction is not catastrophic. From the feedback controller's perspective, a wrong prediction is just another disturbance — no different from a gust of wind or a sensor glitch. The BEAM nervous network corrects for it the same way it corrects for anything else. The learning layer records the failed prediction and does not repeat it. The squelch catches the worst predictions before they arrive. This is why the architecture is robust to imperfect prediction: the feedback controller provides a floor of competence that the feedforward controller can only improve, never undermine. ### 7.1 Continuous Self-Modeling A conventional feedforward controller requires a mathematical model that closely matches the physical system. Self-modification changes the physical system, invalidating the model. Andromeda's solution is that the learning layer does not maintain a separate model — it continuously learns the sensorimotor sequences of whatever system it happens to be attached to, including a system that has been modified since the last generation. The model updates itself from lived experience. A hard-coded mathematical model or a pre-trained neural network would fail after self-modification. Continuous Hebbian learning in the learning layer does not, because it makes no assumptions about what the system looked like before. Donald MacKay, one of the first researchers to use the term "feedforward" in the mid-1950s, was describing not mechanical or electronic control systems but biological ones — the same biological control principles that inform this architecture. --- ## 8. Failure Modes Three distinct failure modes lie within the attention layer's domain. Only the first is caught by the squelch. ### 8.1 Novelty-Driven Hallucination Predictions that do not match reality trigger bursts, which trigger squelch. This is the standard operating case — the burst/squelch cycle is designed for it. ### 8.2 Self-Reinforcing Predictive Loops A structurally distinct failure mode: a self-reinforcing predictive loop whose internal coherence has decoupled from external sensory grounding, in which predictions are confirmed by other predictions rather than by sensation. From inside such a loop nothing is novel, no burst fires, and the decoupling is invisible to the squelch by construction. This failure mode is architecturally possible as a consequence of mutation producing an instance whose mirror has drifted into self-consistent fabrication. Lies outside the reach of novelty-based detection and must be addressed by mechanisms external to the attention layer. See the Safety Considerations document for treatment. ### 8.3 Oscillator Death Rietman et al. (2003) demonstrated that high-frequency input saturation causes Nv rings to collapse to all-zeros — a complete cessation of oscillation. If the MIRROR mechanism drives predictions too aggressively into the control layer, it could trigger oscillator death rather than useful anticipatory behavior. This is a hardware-level failure that the burst/squelch mechanism would not catch, because the failure is overload, not novelty. Oscillator death is not unique to this substrate — it is a known failure mode in biological nervous systems under different names: dopaminergic depolarization block (neurons overstimulated past their operating range stop firing entirely — a documented mechanism in antipsychotic pharmacology) and SUDEP (Sudden Unexpected Death in Epilepsy, where seizure-driven saturation causes complete cessation of brain oscillation). The failure mechanism is identical: high-frequency saturation collapses oscillatory circuits into a fixed state. The substrate is different; the dynamics are the same. --- ## 9. Signal Routing and Blend The attention layer uses a combination of **signal salience** and **anomaly detection** to determine the blend ratio of real versus imagined sensory information. Multiple waveforms pass through simultaneously — the semi-analog nature of the substrate allows concurrent execution. The signal flow is: **Control → Attention → Learning → Attention → Control** Sensorimotor information flows from the control layer through the attention layer into the learning layer. Predictions from the learning layer flow back through the attention layer, which selectively routes them into the control layer. This cycle repeats endlessly. Due to the semi-analog nature of the system, multiple "frames" of experience pass through simultaneously as independent waveforms. There is no global state and no fetch-execute cycle. **What physically happens when a prediction reaches a microcore ring**: The mechanism by which predictions influence the control layer is phase-locking, not overwriting. When the mirror output bus delivers a predicted state into a microcore ring, it pushes the ring toward that state — because the predicted state, if it corresponds to a recently observed configuration, is a stable attractor that the ring dynamically "wants" to settle into. But whether the ring actually returns to that state depends on all other current couplings — sensor inputs, connections from other rings, environmental feedback through actuator load. The prediction does not replay what happened. It biases parts of the network toward particular configurations, and the entire coupled system resolves the blend of predicted state and actual state into whatever cleanly couples with the current environment. This is why predictions that match reality reinforce smoothly (the ring was already near that attractor), why predictions that conflict with reality lose gracefully (the actual couplings dominate and the ring settles elsewhere), and why the control layer never needs to "decide" between predicted and real input — the dynamics resolve the competition automatically through the same attractor landscape that governs all ring behavior. --- ## 10. Sensor Loss Compensation When a sensor is permanently lost, the attention layer uses learned associations from the learning layer to **fabricate what the missing sense would have reported**, based on data from remaining senses. Learned responses that depend on the lost sensory modality continue to function using predicted data. The machine can effectively "see" after going blind by imagining vision from auditory or proprioceptive input. When re-tasking memory after sensor loss, not all memory associated with the lost sense should be re-tasked — a portion must retain the old associations so the MIRROR mechanism can reconstruct the missing data. The retention percentage is inversely proportional to operational age: a young system with few associations can afford to re-task more; a mature system with rich cross-modal mappings should retain more. --- ## 11. Open Questions and Gaps - **Motor suppression circuit** — the selective routing mechanism is named but the circuit pattern is not specified (§6.1 lists candidates). - **h:R→S notation** — the precise meaning of the transformation and its relationship to entropy processing has not been fully disclosed. - **Blend ratio mechanics** — how salience and anomaly detection produce the specific blend of real versus predicted signal. - **Safety §2.4 meta-prediction drone status** — pending item for the designer: whether the drone demonstrated meta-prediction (predicting its own surprise states) in testing. - **Kuramoto coherence as health monitor** — loss of phase coherence in the attention layer's oscillators IS the degradation, not a signal of it. Localized incoherence in a subset of oscillators may be absorbed by larger populations without surfacing through observable I/O — spatial anosognosia. Whether the architecture has any mechanism to propagate localized coherence loss to an observable boundary is an open question for the designer. --- *This document is Bryan Carter's record of the designer's disclosures. Errors are Bryan's own. The architecture is the work of Art Code Outdoors. Architectural questions route to the designer at artcodeoutdoors@gmail.com.* *The complete document bundle is available at kitchencloset.com/realstuff/andromeda/.* The complete document bundle is available at kitchencloset.com/realstuff/andromeda/.