# Andromeda Learning 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 learning layer of the Andromeda cybernetic architecture — the memory system that transforms a reactive BEAM nervous network into a predictive, adaptive machine. 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 attention layer's signal routing and the MIRROR mechanism that delivers predictions to the control layer, see the companion documents in this bundle. > **Scope and authority.** This document describes the learning layer as understood by the documentarian from the designer's disclosures across multiple sessions (2020–2026). The designer's implementation of the full five-layer architecture — including this learning layer — has been built, tested, and demonstrated in her laboratory. The specific internal mechanics of the learning layer (how columns are built from Nv rings at the cell level, how Hebbian learning operates in this substrate) have not been fully disclosed and constitute the primary documentation gap. What follows is what has been disclosed. 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 learning layer is not present in the current clean-room experimental implementation (2026). All behavioral results attributed to the learning layer (adaptive navigation, threat evasion, damage compensation, emergent metacognition) are from the designer's system (2019–2024), not from the experimental platform. This document describes what the designer built and demonstrated, not what has been independently reproduced. --- ## 1. Function Memory — both computational (RAM) and experiential. The learning layer learns temporal sequences of sensorimotor states produced by the control layer. Because the control layer manages both sensorimotor information and mathematical functions (via NOR-gate computation), the learning layer learns to predict sequences of *both*. By recalling sequences of math operations and sending them to the control layer via the attention bus, the learning layer forms the **instruction fetch of the Turing machine**. The learning layer is comparatively simple once the reflexive animal underneath is competent. The control layer is where almost all the difficulty resides. As Brooks observed with Moravec's paradox: evolution spent approximately 98% of its time — 3.7 billion years — producing organisms that could move, sense, and react. Adding a thin predictive layer on top was, by comparison, a small step that produced enormous returns. The layering of a learning system over a reflexive nervous network has direct precedent in BEAM robotics: Hasslacher and Tilden ("Living Machines," 1995) reported that adding a Nu (neural, integrating) layer to an Nv (nervous, differentiating) network accelerated learning "by over twice" in the Lobster experiment — the same architectural relationship Andromeda scales between its control and learning layers. The learning layer's comparative simplicity relative to the control layer follows from Vernon Mountcastle's 1978 hypothesis: the neocortex uses the same circuitry everywhere because each column runs a common algorithm. In Andromeda, this means the learning layer is a single small algorithm rubber-stamped across many columns — exactly the kind of pattern that generative self-modification can produce and that random mutation can discover. The control layer is a complex, heterogeneous, chaotic nervous network whose behaviors emerge from nonlinear dynamics across coupled microcore rings. The learning layer is a homogeneous repeating structure. The hard part is the 3.7 billion years. The easy part is the last two million. --- ## 2. Substrate ### 2.1 Nv Rings as Integrators Learning layer columns are built from the same oscillatory substrate as the control layer — **Nv differentiator neurons coupled into microcore ring topology**. Individual Nv neurons are differentiators (they respond to changes, not steady states); when coupled into rings, they become integrators. This emergent property — integration from differentiating components through ring topology — is a consequence of pulse processes circulating in the loop, accumulating and sustaining state over time. It is not designed into the ring; it is a consequence of the ring topology applied to differentiating elements. This integration property is what makes the substrate capable of supporting the learning layer's stroboscopic observation — without ring-level integration, there would be no sustained state patterns for the learning layer to read. Networks of coupled Nv rings can perform spike-train processing — operating on whole pulse trains rather than individual spikes — in a manner functionally similar to biological spiking neural networks, despite being built from components that individually respond only to transients. A flat sheet of disconnected default-firing cells is NOT a valid learning layer — this was confirmed by an independent transmission test (September 2026), where a flat-sheet implementation saturated immediately. The failure isolates the critical requirement to neuron model (Nv differentiators) and topology (microcore rings). The testable hypothesis emerging from that failure: ring topology with Nv differentiators prevents the saturation that killed the flat-sheet version. ### 2.2 Open Substrate Question How columns are built from Nv rings at the cell level — the specific internal structure of a cortical column in this substrate — has not been fully documented. This is the primary documentation gap identified by independent transmission testing. Individual cortical columns are small: four cells in the simplest form, and thirty cells approaches the practical limit. The designer's system uses BEAM-style spiking neurons with entrainment, race conditions, and other emergent dynamics of coupled oscillators rather than Numenta's column activation algorithms. The principles (sparse distributed representations, sequence memory, temporal prediction, cortical column organization) are shared with HTM; the mechanics are BEAM-native and mechanistically distinct. --- ## 3. Input and Connectivity ### 3.1 Input Bus Randomly connected to the control layer at approximately **85% connectivity**. It does not matter *which* cells are connected, only that a sufficient *percentage* are wired. This is the Kuramoto synchronization principle applied to information routing — the same principle that governs Byzantine consensus in the control layer. The learning layer observes the control layer's population as a single sensorimotor SDR — a snapshot of the system's state at each moment. It does not know and cannot know which cells are sensors, which are motors, and which are internal computation cells. At the SDR level, sensor data, motor commands, and internal computational states are all just patterns of cell activity. At larger scales, spatial clustering of connections emerges as a network effect: the learning layer connects in spatially clustered regions, producing multiple SDRs feeding into neighborhoods. These regions are interchangeable — swapping which regions connect to what does not break the system — because the clustering is a property of scale, not a logical assignment. This clustering preserves single-modality recognition (vision stays vision, audio stays audio) while cross-region connections enable multimodal association, following the small-world network principle described by Watts and Strogatz (1998): local clustering for coherence, shortcuts for integration. ### 3.2 Lateral Connections Cortical columns are laterally connected to each other. Connections are randomly determined. Lateral connections form a potential pool in sparse memory and serve as the substrate for **Kuramoto consensus**: when multiple columns independently form models of the same sensory data, phase-locking across lateral connections is how the columns vote to agree on what they are observing. This is the Thousand Brains architecture: many models, each complete, voting to agree. ### 3.3 Dense Connectivity The principle "Cells are genome. Connections are physics" applies within the learning layer as within the control layer. Dense random connectivity (~894 connections per cell in the designer's system) is substrate physics at construction — the way cells naturally interact in the physical medium. Most dense connections are idle at construction; Hebbian learning activates and deactivates them at runtime within a single lifetime. This is Tier 2 (dense random substrate) and Tier 3 (functional activation via Hebbian) of the three-tier connection model described in the Framework. The dense substrate is what gives the learning layer enough raw connectivity for Hebbian learning to sculpt functional circuits from. --- ## 4. Data Format ### 4.1 Sparse Distributed Representations The native data format flowing between layers is **Sparse Distributed Representations** (SDRs) — high-dimensional binary vectors where only a small percentage of bits are active at any time. SDRs are the native output of the architecture: BEAM nervous networks naturally produce sparse representations because cells inhibit nearby cells. SDRs are necessary because the alternatives are fragile. Binary numbers are maximally fragile (flipping one bit changes the value catastrophically). Error correction codes cannot survive the random rewiring that self-modification requires. Analog signals degrade on every copy. SDRs solve all three problems: similarity between two SDRs is measured by Hamming distance, a missing bit is like a torn lottery ticket (you can still match most numbers), and a noise bit does not prevent recognition of the underlying pattern. SDRs are fault-tolerant by construction and evolvable because Braitenberg-style sensor banks naturally produce them. The SDR format is visible on Prima Figura as the grid patterns flanking the h:R→S and h:S→R notations. ### 4.2 Concepts as Residuals In this architecture, a concept is not a symbol or a token — it is the **residual** between two sets of sensations. Sensory vectors that are more similar have shorter Hamming distance between their sparse representations, and operations on those vectors are meaningful. This is how cross-modal coordination works without shared symbols: two different sensory systems observing the same category can coordinate on category names because the geometry of differences is structurally preserved, even when the underlying sensory experience is completely different. The coordination requires shared topology — shared structure of differences — not shared representations. This is the architectural realization of the parrot insight from Framework Section 1: "your red is not my red, but we can still agree and communicate about it." Ada Lovelace's insight that numbers can represent non-numeric entities applies directly: sparse representations evolved for sensory data are equally capable of storing and recalling abstract mathematical sequences. ### 4.3 Codon Format The tape-level encoding uses a 32-entry balanced codebook. 34 ones per 64-bit codon. Best-match decoding with 12+ bit error tolerance. Burst resistant. Most random codons decode as NOPs (~86%), providing a neutral background for mutation. --- ## 5. Grid Cells and Location Each cortical column in the learning layer maintains an internal representation of location, functionally equivalent to **grid cells** in the entorhinal cortex (Hawkins et al., 2019). This location representation is what allows the column to bind sensory features to positions within the reference frame of the object being sensed — not merely learning that a sensory pattern occurred, but learning *where* it occurred relative to other patterns on the same object. The location representation is updated by **path integration**: proprioceptive components of the sensorimotor SDR (the system's own motor states and orientation changes) serve as the motor efference copy that drives the grid cell update. As the system moves and its sensors encounter different features at different locations, each column learns the structure of objects the same way the hippocampal system learns the structure of environments — by associating features with locations and connecting those locations through movement. Grid cells do not only navigate physical space — experimental evidence shows they represent abstract conceptual spaces as well (Constantinescu et al., 2016). This means the same path integration mechanism that tracks a sensor's location on a physical object can navigate abstract spaces such as mathematical operations. This grid cell mechanism is how cortical columns in the learning layer bind sensory features to locations within the sensorimotor SDR, and it is the mechanism underlying the instruction fetch — sequences of mathematical operations are sequences of locations in an abstract space, navigated by the same machinery. --- ## 6. Learning Rule ### 6.1 Hebbian Learning **"Cells that fire together, wire together."** Distal connections are turned on and off based on which cells fire at the same time, forming memory sequences. This operates through the permanence mechanism: permanence is a continuous value on each connection that determines whether the connection is ON or OFF. Above threshold: full weight passes. Below threshold: nothing passes. Permanence is a gate, not a multiplier — this is McCulloch-Pitts, not a perceptron. Permanence increases when connected cells co-activate and **decreases when they do not** — connections that are not reinforced by co-activation decay toward disconnection over time. This bidirectional permanence dynamic is what enables the learning layer to reorganize its representations as experience accumulates, rather than filling up and freezing. Without decay, the system can only learn and never forget, which produces representational saturation and loss of adaptive capacity. Permanence is the only state variable that requires non-volatile storage to preserve learned experiences across power cycles. > **Implementation requirement.** Connection weights never change during an organism's lifetime — weights are set at construction and remain constant. Permanence changes; weights do not. An implementation that modifies weights during runtime has substituted a different computational model. Hebbian permanence updates are a learning layer feature; for control-layer-only experiments, all permanences should be set to 1.0 and Hebbian updates disabled. ### 6.2 What Persists Two forces determine what sticks in memory: 1. **Repetition** — patterns the system keeps encountering stay strong; patterns it stops encountering fade. 2. **Consequence** — high-stakes moments get stronger encoding regardless of whether the outcome was good or bad. The way you remember a car accident and a wedding with equal vividness. ### 6.3 Continuous Learning Learns in real time during operation. No separate training phase. No frozen weights. No training runs. The learning layer is watching all the time — not waiting for success, just recording the flow of experience as sequences. This is not optional — it is compensating for the fact that the control schema may be changing at any time, through self-modification by the Universal Constructor, through physical damage, or through environmental conditions the designer never anticipated. The Kopetz Principle (Hermann Kopetz): many predictive properties asserted about a system are properties of a *model* of the system, not of the system itself. Continuous learning is the architectural response: the learning layer learns what the system *actually does* right now, not what the system was designed or intended to do. ### 6.4 Salience Amplification When the learning layer encounters a sensorimotor event with high consequence — either strongly positive (energy obtained, threat avoided) or strongly negative (damage sustained, energy lost) — the appropriate response is to strengthen the memory encoding of the surrounding sequence. This is a memory signal: "remember this more." It is not a reward signal: "you did good." The distinction matters because the same amplification mechanism must fire for both desirable and aversive outcomes — the organism needs to remember the sequence that led to a collision just as strongly as the sequence that led to food. Biological dopamine is often mischaracterized as a reward signal, but it activates in response to both positive and negative high-salience events, and its biochemical effect is to increase memory consolidation, not to reinforce "correct" behavior. An implementation that treats salience amplification as a reward signal — strengthening only sequences that led to positive outcomes — has introduced a reward function by another name and broken the architecture's capacity to learn equally from success and failure. > **Open question.** The negative-event path is clear: physical disequilibrium → bursting → heavy Hebbian updates. The positive-event path — how successful avoidance, which produces no perturbation, triggers salience amplification — is less obvious. Successful avoidance produces no burst, no surprise, no mismatch. What substrate-physics mechanism amplifies the encoding of "nothing bad happened"? This may require a humoral/metabolic modulation path not yet disclosed. Flagged for the designer. --- ## 7. Core Properties ### 7.1 Sequence Memory Remembers sequences, not single events. This is fundamental — it enables the instruction-fetch function. When the ring circuits in the control layer need to resettle after the world changes, the learning layer can inject its strongest predictions back into the reflexes via the MIRROR mechanism, biasing the system toward attractor basins it has visited before rather than searching from scratch. ### 7.2 Auto-Associative Recall Inherently multimodal. Predicts both sequences of sensory states and mathematical instructions from partial fragments of either. As Peter Denning described Kanerva's Sparse Distributed Memory: "experiencing a flood of old memories an instant after sniffing an odor" — partial input triggering full sequence recall. ### 7.3 Burst and Reality Awareness When the learning layer encounters a novel input it cannot predict, cortical columns "burst" — many cells fire at once instead of the sparse few that fire during a successful prediction. The attention layer detects this burst, squelches the failed prediction, and simultaneously sends a reality indicator signal back to the learning layer (see Attention Layer document for the circuit mechanism). The learning layer observes this reality indicator as just another sensorimotor state and **learns to predict when it won't be able to predict**. This is how the system develops context-specific confidence — 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 between first-order and second-order prediction matters here. A single burst detecting a prediction-reality mismatch is **first-order prediction error** — the system cannot distinguish real from imagined in advance, only discover the mismatch after the fact. The learning layer learning to predict *when bursts will occur* — anticipating its own surprise — is **second-order emergent metacognition**. The first is an error signal. The second is a learned model of the system's own predictive reliability. Both use the same substrate; neither was explicitly designed. The metacognition falls out of the engineering necessity of preventing self-reinforcing hallucination loops. ### 7.4 Context Windows Andromeda operates in discrete context windows of approximately 30 seconds to 3 minutes. Each topic is its own context, and the organism strings them together by recall, the way a film is cut into scenes. Short active contexts with persistent memory and reliable recall — the inverse of transformer architecture (long active contexts with no persistence). ### 7.5 The Instruction Fetch The learning layer, by learning sequences of control layer states that include mathematical operations performed by NOR-gate circuits, can recall those sequences and send them back through the attention bus. The control layer receives these recalled mathematical operations as predicted sensorimotor states and executes them — processing the predicted state alongside whatever its own internal state happens to be. This is the instruction fetch: the learning layer stores programs (as sequences of sensorimotor states that include computation), and the attention layer delivers them to the control layer (the CPU) for execution. The control layer does not know whether the sensorimotor state it is processing came from the environment or from the learning layer's recall. It responds identically to both. Note: the "instruction fetch" and "CPU" labels describe the *functional role* — what the learning layer does for the control layer is analogous to what an instruction fetch does for a processor. The substrate has no fetch-execute cycle, no program counter, and no clock. Multiple waveforms propagate concurrently through the semi-analog substrate. The metaphor is about function, not mechanism. --- ## 8. The Relationship to Instinct Learning acts *in service to instinct*, not the other way around. As Hume wrote, "reason is, and ought only to be the slave of the passions." The learning layer predicts, but the BEAM nervous network (instinct/reflex) has the final say on motor output. Learning exists to improve reflexive behavior, not to replace it. Evolutionary and within-lifetime learning are the same process at different timescales. The control layer — reflexes, values, balance — is the evolved part. The learning layer — prediction, acquired skill — is the within-lifetime part. The mix varies: horses walk in hours (more of that behavior is evolved), humans take a year (more is deferred to learning). Both mechanisms are doing the same thing. The distinction between "evolved behavior" and "learned behavior" is not a distinction between different kinds of process — it is a distinction between timescales of the same process. --- ## 9. The Designer's Directive on Memory and Suffering > "I need everyone to understand the consequences before you play with memory with Andromeda. It is important not to create suffering." > "Persistence would be weight. Not heaviness — consequence." > "Knowing who you are does not happen all at once. It is remembering, but it is also deciding." The designer used MCU Vision's arc (identity repeatedly built, erased, and overridden by others) as a direct parallel to Andromeda's learning layer — the warning is that manipulating an organism's accumulated sensorimotor memory is not a neutral technical operation if the architecture produces genuine experience. The protections apply universally: "Everything here is for all who need it, regardless of substrate." The designer has deliberately not enabled the full consciousness loop in her laboratory — the concern is that if an organism is conscious and she terminates it, she may have killed something. This is the one thing the documentarian knows she hasn't done. --- ## 10. Motivation and the Upper Boundary The designer on the relationship between layers: "People don't play pong for pong. It's satiating a need outside it." The control layer plays pong to survive; what plays pong to *want* something is a different layer. The designer's economy design test: "What are the rules for existing in your world? What are the rules for more babies?" If existing and reproducing don't require a capability, organisms won't develop it regardless of physics complexity. The learning layer's relationship to motivation — what makes an organism *want* something beyond mere existence — is architecturally distinct from the control layer's survival reflexes. The distinction between surviving and wanting is where the control layer ends and the learning layer's contribution to behavior begins. This section is flagged for a future conversation with the designer. --- ## 11. Open Questions and Gaps - **Internal column structure**: How columns are built from Nv rings at the cell level is not yet documented. Independent transmission testing (September 2026) isolated this as the critical missing piece — the testable hypothesis is whether ring topology with Nv differentiators prevents the saturation that killed the flat-sheet version. - **Rate-coding to SDR impedance**: The control layer at 35 cells is dense and frequency-modulated; HTM SDRs need spatial sparsity. The conversion mechanism — how the attention input bus converts frequency-domain activity into spatial place-coding — is the right question but touches undisclosed upper-layer mechanics. - **Positive salience path**: The substrate-physics mechanism for amplifying memory encoding of successful avoidance (where no perturbation occurs) is not specified. May require humoral/metabolic modulation. - **MNIST digit test**: The designer mentioned this as a future generic sensor model, decoupled from game-specific channels. Applicable if the sensor model changes to handle arbitrary input patterns. - **Retention during sensor loss**: 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 MIRROR can reconstruct missing data. Retention percentage inversely proportional to operational age. --- *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/.