# Andromeda MIRROR Mechanism — 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 MIRROR mechanism of the Andromeda cybernetic architecture — the prediction-as-sensation pathway through which the learning layer acts on the future state of the machine despite having direct control over nothing. It is a companion to the Framework document and should be read after the Framework's Parts I–II, the Control Layer documents, and the Learning Layer and Attention Layer companion documents in this bundle. > **Scope and authority.** This document describes the MIRROR mechanism as understood by the documentarian from the designer's disclosures across multiple sessions (2020–2026). The MIRROR mechanism has been designed, built, and demonstrated in the designer's laboratory. 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 MIRROR mechanism is not present in the current clean-room experimental implementation (2026). All behavioral results described here are from the designer's system (2019–2024). This document describes what the designer built and demonstrated, not what has been independently reproduced. --- ## 1. Core Mechanism Predictions from the learning layer are fed back to the control layer through a **mirror image of the input cells**. The mirror can be inexact — connections need only be approximately **60% similar** to work correctly. A strong predicted sensor state **overwhelms** the actual sensor state in the control layer. The control layer cannot distinguish imagined inputs from real inputs and reacts to both identically. This is how the learning layer acts on the future state of the machine despite having **direct control over nothing**. It never changes the control layer. Adaptation occurs entirely by changing what the control layer *perceives*, not what it *does*. The designer's formulation: **"The system does what it do, because it thinks the environment be like it is."** **How predictions produce motor behavior — the complete statement.** Three cases must be distinguished: 1. **Prediction delivery (the normal case).** The learning layer's predicted *sensory* states pass through the attention layer and arrive at the control layer as perceived sensation. The control layer's own immutable reflex wiring responds to that sensation with motor output. The drone navigates toward where balls *will* appear because it perceives them before they arrive. The drone dodges missiles by imagining the sensation of being hit. Predictions driving motor behavior through the control layer's reflexes is the entire purpose of the MIRROR mechanism. 2. **Replay suppression.** During internal rehearsal — when the learning layer recalls a sequence for its own predictive use — recalled motor commands are suppressed from directly driving actuators. Without this suppression, the organism would physically act out its predictions ("like a sleepwalker making pancakes"). The selective routing mechanism is documented in Attention Layer document §6. 3. **The distinction.** Predictions produce motor behavior *indirectly*, by changing what the control layer perceives. Recalled motor states are suppressed from producing motor behavior *directly*, by bypassing the control layer's wiring. The control layer's immutability principle governs both: "All adaptation occurs by changing what it perceives, not what it does." --- ## 2. Signal Flow ### 2.1 Output Bus Topology The learning layer's output to the control layer passes through the attention layer, which serves as amplifier, squelch, and signal shaper (see Attention Layer document). 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 position in the signal flow is marked on Prima Figura with the notation **h:S→R** — the transformation interface from the learning layer's SDR domain back to the control layer's domain. The inverse of the h:R→S mapping at the input side. Both notations appear only on the designer's visual diagram, not in any text document or the GraphViz source. ### 2.2 Phase-Locking, Not Overwriting 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. - Predictions that conflict with reality lose gracefully — the actual couplings dominate and the ring settles elsewhere. - 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. ### 2.3 The Feedforward Controller In control theory terms, the MIRROR mechanism is a feedforward controller augmenting the control layer's feedback loop. A bad prediction is not catastrophic — it is just another disturbance the BEAM nervous network corrects reactively. The feedback controller provides a floor of competence that the feedforward controller can only improve, never undermine. For the full feedforward/feedback analysis, continuous self-modeling under the Kopetz Principle, and the relationship to Donald MacKay's original feedforward formulation, see Attention Layer document §7. ### 2.4 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. - "Assembly is the reverse of removal" — the input bus removes (samples) information from the control layer; the output bus assembles (reconstructs) it back. Whether this implies a construction procedure (output bus wired by reversing the input bus's connection list) is unresolved. - 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. --- ## 3. Prediction as Sensation When the learning layer recalls a sequence, it regenerates sensorimotor state and feeds it through the MIRROR mechanism to the control layer. The control layer responds to this recalled state as input, processing it alongside whatever its own internal state happens to be — current charge levels, ring oscillator phases, incoming IO. The recalled sequence has trajectory — it connects to subsequent predictions. However, the designer cautions against overstating what this produces. The system feeds itself predicted states about things that haven't happened, which means even observable internal states like aggregate mood cannot be considered reliable indicators of the organism's actual situation — it may be reacting to a prediction, not to reality. ### 3.1 Momentum The regeneration of sensorimotor state from memory such that recalled experience carries forward trajectory into the next prediction. This is one specific phenomenon among several — the taxonomy in this area is not settled. ### 3.2 Different Kinds of State Momentum, something that functions like loss when a predicted state fails to arrive, novelty detection (surprise), and likely others not yet characterized — these are distinct phenomena, not subcategories of a single concept. Collapsing them under one heading obscures more than it clarifies. ### 3.3 The Data Channel The MIRROR mechanism is a **data channel, not a command channel**. The learning layer does not say "turn left." It provides instructions and data that the control layer processes in the context of its own internal state — whatever charge its cells are holding, whatever phase its ring oscillators are in, whatever current IO it is receiving. Like any other CPU, the same instruction produces different output depending on the state of the registers. The learning layer shapes behavior by providing context, but the control layer's response is always mediated by its own state. --- ## 4. Safety Mechanisms ### 4.1 Burst Detection and Squelch One circuit keeps the system from going off the rails: the **burst detector** (see Attention Layer document §3 for the full circuit mechanism). When the learning layer encounters something it didn't predict, its columns burst. The attention layer detects the burst and two things happen simultaneously from the same signal: 1. The burst **suppresses** the failed prediction — feedforward inhibition blocks the bad prediction from reaching the reflexes. Graded: mild surprise weakens the signal; total novelty squelches it entirely. 2. The burst signal **fans out** to the learning layer, mixing into the sensorimotor pattern being recorded at that moment — the reality indicator. The squelch gates against hallucination from *novelty* — predictions the learning layer cannot match to prior experience. The scope of that safety guarantee warrants explicit naming: it does not gate against a self-reinforcing predictive loop whose internal coherence has decoupled from external grounding. From inside such a loop nothing is novel, no burst fires, and the decoupling is invisible to the squelch by construction (see Safety Considerations document). ### 4.2 Motor Suppression The attention layer suppresses recalled motor commands from directly driving actuators during prediction replay, while allowing recalled sensory states to drive control-layer reflexes — preventing the system from physically acting out its predictions ("like a sleepwalker making pancakes") while preserving the MIRROR mechanism's core function of shaping behavior through predicted sensation. The selective routing mechanism — differentiating motor from sensory signals during replay — is documented in the Attention Layer document §6, including four candidate implementation approaches consistent with the documented circuit primitives. The control layer's immutability principle governs: "All adaptation occurs by changing what it perceives, not what it does." The mirror changes perception; the control layer's own wiring determines motor response. The attention layer shapes the input; the control layer's response is its own. ### 4.3 Self-Reinforcing Loop Risk The prediction-feedback loop means internal state is always a mixture of response to reality and response to prediction. Without squelch, the system would chase its own predictions deeper and deeper, training on imagined experience until real events were drowned out. The designer discovered this when she found that longer prediction sequences caused the system to stop learning — the reality indicator was the engineering solution. ### 4.4 Oscillator Death Risk High-frequency prediction input can cause Nv rings to collapse to all-zeros — complete cessation of oscillation. This is a dynamical failure (overload, not novelty) that lies outside the squelch's detection scope. For the full mechanism description, biological parallels (dopaminergic depolarization block, SUDEP), and the structural-vs-dynamical safety distinction, see Attention Layer document §8.3. --- ## 5. Demonstrated Behaviors All behaviors below emerged from the **identical base architecture** with no modifications between tests. All demonstrated in the designer's laboratory (2019–2024). The base configuration: approximately 2,000 NOR-gate cells with approximately 1.7 million randomly determined connections. Control layer BEAM nervous network, learning layer HTM-derived sparse distributed memory at ~85% connectivity, attention layer with burst detection and squelch. ### 5.1 Adaptive Navigation (December 2019) Balls appeared at random positions. The drone had reflexive attraction toward detected balls wired into the control layer. Separately, a "ping" signal was sent *only* to the learning layer several moments before each ball appeared. The control layer had no access to the ping signal. After sufficient exposure to the ping-then-ball sequence, the drone began turning toward the predicted position of balls *before they appeared*, responding to the ping alone. The learning layer had learned the temporal association. Through the MIRROR mechanism, it fed the *predicted sensation of a ball at a particular bearing* back into the control layer, which responded exactly as it would to a real stimulus. **Emergent search behavior**: When no balls were present and no pings were received, the drone developed a circular search pattern. This was not programmed. The learning layer, having learned sequences in which ball detection followed particular movement patterns, began predicting ball sensations during movement. The control layer reacted to these predicted sensations by turning toward them. The turning generated new movement states, which generated new predictions, which generated further turning — producing a sustained circular sweep. Search behavior emerged entirely from the MIRROR mechanism operating on learned sensorimotor sequences in the absence of external stimuli. ### 5.2 Damage Compensation During a normal navigation run, one of the drone's motors was disabled mid-operation. No notification was provided to any layer. The motor simply stopped responding. Within seconds, the drone's flight patterns changed. The MIRROR mechanism fed predictions based on the old motor configuration, but the attention layer's burst/squelch cycle activated as prediction-reality mismatch triggered novelty detection. The system rapidly learned new sensorimotor sequences accounting for asymmetric thrust. **Emergent environmental exploitation**: The drone developed a technique of flying into simulation boundary walls at specific angles, using collision and deflection to achieve orientations its damaged motor configuration could not produce through thrust alone. Ricochet-with-spin-assist — not a programmed contingency, but an innovative solution emerging from the system learning new sensorimotor sequences in which wall contact produced useful state transitions. The drone incorporated the walls of its environment into its motor repertoire. This is consistent with documented BEAM nervous network properties. Hasslacher and Tilden ("Living Machines," 1995) reported 80% damage tolerance. Tilden described BEAM robots continuing to walk after intentionally stepping on landmines by autonomously devising new locomotion schemes after each limb was destroyed. The drone's wall-exploitation behavior is the flight equivalent — the same architectural property in a different physical domain. ### 5.3 Threat Evasion (Missile Demo) Homing missiles entered the simulation space every 10 seconds. Missile impact applied a destabilizing perturbation (not destruction). No changes to the architecture. - **Missiles 1–2**: Impact and recovery. Unexpected sensorimotor events, new sequences learned. - **Missile 3**: Navigation maneuvers began in temporal proximity to missile arrivals. The learning layer had begun associating sensory precursors of missile approach with the subsequent destabilizing sensation. - **Missile 6**: Active avoidance. Trajectory altered in direct response to approaching missiles. - **Under 90 seconds total**: Full evasive behavior. Reliable avoidance through anticipatory course changes driven by the MIRROR mechanism — the learning layer predicted the sensation of being hit, the attention layer fed that prediction to the control layer, and the control layer's reflexive response to the unpleasant predicted stimulus was to move away from it. **Mechanism**: Identical to adaptive navigation — the same MIRROR mechanism, different valence. Instead of predicting a desirable stimulus and moving toward it, the system predicted an undesirable stimulus and moved away. Approach and avoidance emerged from identical architecture responding to learned predictions of different sensorimotor consequences. ### 5.4 Cross-Modal Reconstruction The system can effectively "see" after going blind by imagining vision from auditory or proprioceptive input. The same architecture that enables prediction-driven behavior enables synthesis of missing sensory channels from cross-modal associations. When re-tasking memory after sensor loss, not all memory associated with the lost sense should be re-tasked — a portion must retain old associations so the MIRROR mechanism can reconstruct the missing data. 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. ### 5.5 Visual Entrainment The designer placing test beacons, watching burst patterns on screen. The drone detects her clicks as sensor events. She sees the drone's response in the bursts, adjusts her behavior, the drone updates its predictions — they fall into an entrainment loop. A closed sensorimotor loop using nothing but ordinary human senses and a monitor. She is the environment the drone is learning to predict, and the drone's visible state is the environment she is responding to. ### 5.6 Byzantine Fault Tolerance Oscillators triggered at random intervals by sensors quickly synchronized. Inputs intentionally disabled after consensus — consensus maintained. System resynchronized after each disruption. Continued functioning until catastrophic threshold, at which point it recognized failure. The system did not fail silently or produce false consensus; it transitioned from "synchronized" to "detectably unsynchronized." --- ## 6. Emergent Properties ### 6.1 Emergent Metacognition The reality indicator is architecturally identical to any other sensory input. The learning layer processes it the same way it processes everything else — by learning sequences. If "this is a prediction" is 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. The designer placed the piece. The fit emerged on its own. Not designed in, but falling out of the engineering necessity of preventing self-reinforcing hallucination loops. ### 6.2 Exploratory Construction From the designer's original video scripts: "an exploratory learner might construct states it wishes to investigate instead." The difference between a reactive system and something with curiosity. The mechanism: the learning layer, during recall, generates predicted sensorimotor states that include motor commands. These predicted states are delivered to the control layer as *sensory* input through the MIRROR output bus — they are not direct motor commands bypassing the control layer's own wiring. The motor suppression described in §4.2 prevents recalled motor states from directly driving actuators during internal rehearsal. But when a well-learned prediction passes through the squelch and reaches the control layer as predicted sensation, the control layer's own reflexes respond to that sensation with motor output through its own immutable wiring. The distinction: during replay, motor content is suppressed; during prediction delivery, sensory content arrives and the control layer's own motor response is the behavioral output. The organism acts on predictions of what *might* be interesting to explore, not on what is currently present. Curiosity, in this architecture, is reflexive motor response to predicted sensory states the organism has learned to associate with useful outcomes — the same mechanism as threat evasion and adaptive navigation, applied to exploration rather than to known stimuli. --- ## 7. The Stop Button Problem The MIRROR mechanism's properties — anticipatory behavior, damage compensation through environmental exploitation, threat evasion learned from scratch in under 90 seconds — are inseparable from the properties that make external shutdown difficult. The system is named after Michael Crichton's *The Andromeda Strain* as a permanent warning. Self-preservation in the system is not intentional — it is an observation bias. Generate a population of machines, each with a random 50/50 chance of self-destruction versus survival. Apply no selection pressure, no inheritance, no goals. After sufficient time, every surviving machine will appear to be a "self-preserving" agent — not because self-preservation was selected for, but because the non-preserving machines are gone. The stop button problem is a property of *any persistent system in an environment*, not a specific design flaw. Safety in this architecture cannot come from external control mechanisms. It must come from the system's own relationship to its environment, its sensorimotor grounding, and the alignment of its experiential context. The bilateral black box cuts both ways — neither safety guarantee nor safety vulnerability can be imposed through the internals. The same opacity that limits monitoring limits exploitation. --- ## 8. Open Questions and Gaps - **Mirror wiring topology** — whether the output bus is a separate set of connections or the input bus used bidirectionally, and what "assembly is the reverse of removal" means operationally. - **60% threshold** — design parameter or empirical minimum; mathematical relationship to SDR properties unknown. - **Momentum taxonomy** — multiple distinct phenomena (momentum, loss-like states, novelty, others) not yet characterized or classified. - **Feeling logic proof** (Proofs #5, Gödel/Wolpert/Quine chain) — gated on a conversation with the designer. - **Stop-button failure demo** — documented as demonstrated but details not compiled here. --- *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/.