A hierarchical lattice memory — three layers of grammar-constrained graphs that learn transition structure from text. Not a neural network, not a database. A topology that encodes statistical regularity as bond strength and neighbor structure. The hexagram rule table + ADSR envelope + Wu Xing cycle give it dynamics. The enzymes give it digestion. The reflex arcs give it fast paths.
What exists: Three-layer lattice (char 16.5M → morpheme 500K → word 100K). Spreading activation engine. Hexagram 64-rule table for node behavior. Reflex arcs (named deliberately). Hierarchical inference threads up, activates at top, projects down. ADSR envelope modulates node lifecycle.
Cognate to: Sensory cortex layering (V1→V2→V4). Reflex arcs = spinal reflexes (fast, deterministic, bypass higher cognition). Spreading activation = action potential propagation. ADSR = neuronal refractory period. The hexagram rules are the ion channel logic — local state determines behavior.
Missing: No attention mechanism (selective amplification of specific activation regions). No working memory (transient buffer that persists across inference calls). No long-term potentiation (the lattice doesn't strengthen bonds during inference, only during training). No inhibitory neurons — all activation is excitatory, there's no lateral inhibition to sharpen responses.
What exists: Five-phase cycle modulates global parameters. Wood expands, Fire generates, Earth stabilizes, Metal prunes, Water clears. Ke corrections handle failures (growth corrects stagnation, testing corrects excess).
Cognate to: Hormonal regulation. Each phase is like a dominant hormone state — cortisol (Wood/expansion under stress), adrenaline (Fire/rapid generation), serotonin (Earth/stabilization), testosterone (Metal/aggressive pruning), melatonin (Water/reset and preparation).
Missing: No feedback loops between phases and actual lattice health. The governor runs a fixed cycle — a real endocrine system would detect cortisol levels and modulate accordingly. No circadian rhythm (periodic maintenance independent of query load). No stress response (acute adaptation to adversarial input).
What exists: Rune enzyme classifies characters. Morpheme enzyme decomposes text via arc matching. Word enzyme cascades through two arc levels. grow.Run does BFS bonding — the actual absorption. feedCorpus reads files, digests them, bonds elements to the lattice. dissolveWeak removes poorly-bonded occupants.
Cognate to: Enzymes literally named after biology. Text → enzyme → elements = food → enzymes → nutrients. grow.Run bonding = nutrient absorption into tissue. dissolveWeak = excretion of waste. The three enzyme types are like amylase (carbs/chars), protease (proteins/morphemes), lipase (fats/words) — each breaks down at a different structural level.
Missing: No gut flora (symbiotic pre-processing — could be external filters or transformers that condition text before enzyme decomposition). No hunger signal (the system doesn't know when it needs more training data). No satiety (it will eat until you stop feeding it, even if the lattice is saturated). No bile/emulsification (handling of adversarial or malformed input before digestion).
What exists: Grammar defines node types (tags) and constraint functions. The lattice graph structure (neighbor topology, block organization) is rigid once built. BuildGrammarLattice creates the skeleton with typed nodes.
Cognate to: Grammar constraints = bone shape constraining how joints move. The fixed topology = skeletal frame. Block organization = vertebrae. Node types = different bone densities (compact vs spongy).
Missing: No cartilage (flexible boundary zones between grammar regions). No bone remodeling (the grammar is fixed at creation — a living skeleton remodels under stress). No growth plates (the lattice doesn't grow new nodes after initial creation, only occupies existing ones).
What exists: Reflex arcs are deterministic paths — the fast, reliable outputs. The activation engine drives them. ArcIndex.MatchAll does greedy longest-match decomposition.
Cognate to: Reflex arcs = skeletal muscles attached to bones (grammar). Lock-in = muscle tone. The dominance threshold for arc extension = motor recruitment — only fire if the signal is strong enough. Greedy matching = ballistic movement (commit to the longest match, don't second-guess).
Missing: No muscle memory (frequently-used arcs should get faster/cheaper). No fatigue (arcs don't degrade with overuse). No antagonist muscles (opposing arcs that prevent over-extension). No smooth muscle (involuntary background processing).
What exists: dissolveWeak is primitive innate immunity — remove anything below a threshold. The dominance ratio in arc extraction is selectivity (reject ambiguous patterns).
Missing: Almost everything. No pathogen detection (identifying adversarial or corrupted input). No adaptive immunity (learning to recognize and reject specific attack patterns). No inflammation response (local activation surge around damaged/corrupted lattice regions). No memory B-cells (remembering past attacks). For a Nostr sentry, this is critical — the immune system IS the detection layer.
What's missing: There's no persistent flow of information through the system independent of queries. In biology, blood continuously circulates, carrying oxygen, nutrients, hormones, and waste products. The lattice has no equivalent — activation only happens when triggered by inference. A circulatory system would be a background process that periodically propagates signals through the lattice, maintaining baseline activation, distributing Wu Xing phase signals globally, and carrying metabolic byproducts (dissolved weak bonds) to elimination.
What's missing: The I/O cycle. Breathing is continuous, rhythmic intake and output. The system currently does batch processing — eat a corpus, train, then infer on demand. A respiratory system would be continuous streaming: inhale (read from Nostr relay firehose), process (activate/bond), exhale (emit verdicts/responses). The sentry (cmd/sentry) is a primitive lung — it breathes Nostr events — but it's disconnected from the lattice hierarchy.
What's missing: The ability to spawn new lattice instances, fork a trained lattice into variants, share learned structure between instances. Mitosis (splitting a lattice into two specialized copies). Meiosis (combining structure from two differently-trained lattices). Gametes (portable arc databases that can seed a new lattice — actually, the .arcs.bin files are half of this already).
What's missing: The boundary membrane. What comes in, what goes out, what's filtered. Input validation, encoding normalization, rate limiting. The skin is the first line of defense and the primary sensory surface. Currently text goes straight to the enzyme with no boundary processing.
What's missing: Drainage and cleanup. dissolveWeak is a start, but there's no continuous lymphatic flow that identifies and removes dead/orphaned nodes, reclaims unused lattice capacity, or transports immune signals to where they're needed.
What's missing: Self-awareness of internal state. The lattice has Health() (occupancy rate) but no continuous self-monitoring. A proprioceptive system would track: activation distribution, bond strength histogram, phase distribution across Wu Xing, lattice temperature (rate of change), and feed this back into governance.
PRESENT (6/12):
Nervous lattice + activation + hexagram + reflex arcs
Endocrine Wu Xing governance + ADSR envelope
Digestive enzymes + grow.Run + dissolveWeak
Skeletal grammar + lattice topology
Muscular reflex arcs + activation engine
Immune dissolveWeak (primitive innate only)
ABSENT (6/12):
Circulatory background signal flow
Respiratory continuous I/O streaming
Reproductive lattice forking/combining
Integumentary boundary membrane / input filtering
Lymphatic continuous cleanup / dead node drainage
Proprioceptive self-monitoring / internal state awareness
The most impactful missing pieces, in order: