Status: pre-launch. This page describes target design; this document does not establish a deployed service.
Skyzai.ORG — Specialized vs General
Current reading — 2026-09-27
Specialized routing is a testable strategy to compare with task-specific alternatives. Biology does not prove universal superiority, exact architecture or measured compute savings. Latency and percentage claims below need their own evidence. [I]
This is an editorial boundary, not a new runtime contract, upstream repeal or proof of adoption. See current status and the product roles. Original revision dates remain as recorded below.
Historical research draft — retained on 2026-09-27; not current terms or implementation instructions
200 cell types > one super-cell. Three universal layers. McGilchrist routing. Why specialization without coordination is just a pile of cells.
The Biological Proof
The human body has approximately 200 distinct cell types. Neurons, hepatocytes, osteocytes, cardiomyocytes, erythrocytes, lymphocytes. Each is optimized for exactly one function. A neuron cannot digest food. A liver cell cannot transmit electrical impulses. A red blood cell sacrificed its nucleus to carry more hemoglobin. [E]
No cell does everything. Every cell does one thing well. The organism emerges from their coordination, not from any cell's individual capability. [I]
This is not a design choice. It is the outcome of 3.8 billion years of selection pressure. Every organism that attempted generalist cells — cells that could sort-of-digest and sort-of-signal and sort-of-defend — was outcompeted by organisms with specialists. [E]
The Four Coordination Primitives
Specialization without coordination is just a pile of cells. The advantage comes from the interface layer — and biology developed exactly four: [S]
| Biological Primitive | Skyzai Equivalent | Type | What It Does |
|---|---|---|---|
| Hormones (broadcast) | RELAY | 1:N | One gland emits; every cell with the right receptor responds |
| Nerves (point-to-point) | FLOW | 1:1 | Axons carry specific signals between specific regions |
| Immune signalling (mesh) | MESH (over SPECTRE nodes) | N:N | Cytokines propagate through tissue, recruiting responses from multiple cell types |
| Convergence (fusion) | AXIOM | N:1 | Multiple sensory inputs fuse in cortical integration areas |
These four primitives are the organism. Without them, you have a tumor — cells multiplying without purpose. [S]
The Three Universal Layers
Specialization happens within a three-layer architecture that is sensor-agnostic, compute-agnostic, and actuator-agnostic: [S]
Layer 1: SENSORS → Circle (IS) — Every sensor. Any modality.
Layer 2: COMPUTE → Rosetta (COULD) — Every model. Any architecture.
Layer 3: ACTUATORS → Skyzai execution — Every wallet, robot, API. Any effector.
Each layer is agnostic to the implementation details of the layers above and below it. A webcam and a Bloomberg terminal both belong to Layer 1. A BitNet ternary model and a GPT-class transformer both belong to Layer 2. A Hedera wallet and a robotic arm both belong to Layer 3. This is layer grammar, not a claim that every example is already live in today's runtime. [S]
Layer Interface Contracts
| Layer | Input | Output | Invariant |
|---|---|---|---|
| Sensors | Raw data (any modality) | Observation { source_id, timestamp, claim, evidence_tier, signature } | Verified, not raw |
| Compute | Observation[] | Prediction { observation_ids[], probability, confidence_interval, model_class } | Probabilistic, not asserted |
| Actuators | Prediction + decision | Action { instruction, authorization_chain[], k0_receipt_anchor, signature } | Signed, not assumed; external anchoring depends on lane/runtime proof |
The agnosticism at each layer means the organism can evolve its sensory apparatus, its cognitive capacity, and its motor capabilities independently. Neurons can be upgraded without rebuilding the gut. Eyes can be replaced without rewiring the brain. This is how biological organisms achieve longevity: modular replacement within a stable architecture. [I]
McGilchrist Routing — The Right Model for the Right Job
Iain McGilchrist's hemispheric hypothesis maps directly onto model routing. The left hemisphere (focused, narrow, already-known) handles routine tasks cheaply. The right hemisphere (broad, novel, contextual) handles genuinely new situations with capable models. The corpus callosum bridges them at decision points. [I]
| Lane | Hemisphere | L-Levels | Models | Cost | When |
|---|---|---|---|---|---|
| LEFT | Left (focused) | L1, L3, L6 | BitNet on-device / Haiku / Sonnet | $0-Low | Known patterns, routine processing, compression |
| RIGHT | Right (broad) | L2, L5 | Opus / Gemini 1M | $$-$$$ | Novel situations, exploration, long-context design |
| CALLOSUM | Bridge | L4, L7 | Opus at specific temperatures | $$-$$$$ | Decision points, constitutional moments |
Routing logic: [S]
- Every signal enters LEFT (L1 on-device — cheap, fast, 29ms)
- If L1 detects novelty or anomaly, escalate to RIGHT (L2 exploration)
- L3 (LEFT again) filters and ranks the explorations deductively
- L4 (CALLOSUM) makes the decision
- If L4 is blocked, escalate through L5 (RIGHT), L6 (LEFT), L7 (CALLOSUM)
Most signals never leave LEFT. The organism processes 95% of inputs on-device for zero cost. Only genuinely novel or important signals reach cloud models. This is how the organism scales: the autonomic nervous system handles volume while the cortex handles complexity. [S]
The Pipeline Enforces Specialization
The action-control projection (IS → COULD → SHOULD → AUTHORIZE (Natural Person's own typed private act; Legal Person's lawful external authority plus required executors; or the signer(s) required by applicable law, the entity's governing instrument, or a valid competent mandate bind the exact public-DAV consequence in a complete valid bound PRISM decision receipt; PRISM records and verifies only; AI and agents stage unsigned proposals only and never execute a public-DAV consequence) → EXECUTE UNDER EXACT TYPED AUTHORITY (for the public-DAV branch, deterministic execution is limited to the exact consequence bound in that complete valid bound PRISM decision receipt) → RECEIPT → REVIEW) prevents the contamination that occurs when conjugate functions merge into one model or when cognition is treated as permission: [S]
| Failure Mode | What Happens | Example |
|---|---|---|
| IS contaminated by SHOULD | Hallucinated observations — seeing what you want to see | A model that wants to trade "sees" bullish signals |
| COULD contaminated by IS | Anchored predictions — predicting what you already observed | A model that sees rising prices predicts rising prices |
| SHOULD contaminated by COULD | Rationalized decisions — deciding what you already predicted | A model that predicts X decides X without deliberation |
| ACT bypasses SHOULD | Impulsive execution — acting before deliberation completes | A model executes on a prediction that was never adjudicated |
These are not bugs. They are the inevitable consequence of merging conjugate functions into a single optimization target. The three-layer boundary is a constitutional firewall. [S]
The Medical Model — A Hospital Is Not One Doctor
A hospital does not employ one type of doctor. It employs cardiologists, neurologists, radiologists, anesthesiologists, pathologists — each trained for a decade in one domain. A radiologist who also performs surgery is a liability. A surgeon who also reads MRIs misses findings. [I]
The hospital's value is not in any single specialist. It is in the referral pathway:
GP observes symptoms (IS)
→ Routes to specialist for diagnosis (COULD)
→ Recommends surgical plan (SHOULD)
→ Surgeon operates (ACT)
The pipeline is the clinical pathway. [S]
The Financial Model — A Fund Is Not One Manager
Asset management followed the same trajectory. The 1960s general-purpose fund manager — one person picking stocks, bonds, and currencies — was replaced by a pipeline of specialists: [I]
| Role | Stage | Function |
|---|---|---|
| Quantitative analysts | IS | Process market data. Signal extraction. Pattern recognition. |
| Risk modelers | COULD | Compute probability distributions. Stress-test portfolios. |
| Portfolio managers | SHOULD | Allocate capital given risk constraints. Make trade decisions. |
| Execution desks | ACT | Route orders. Minimize slippage. Settle trades. |
A single generalist model cannot match this pipeline. Not because it lacks intelligence, but because the functions are conjugate — optimizing one degrades another. [S]
The Principle
| Domain | Specialized Components | Coordination Layer | Result |
|---|---|---|---|
| Biology | 200 cell types | Nervous/circulatory/immune/endocrine systems | Living organism |
| Medicine | Medical specialties | Clinical referral pathways | Hospital care |
| Finance | Quants/risks/PMs/traders | Investment process | Fund performance |
| Skyzai | AI models (any provider) | Rosetta + the pipeline + SPECTRE | Organism intelligence |
Specialization is not a luxury. It is a thermodynamic necessity. Organisms that specialize their subsystems and invest in coordination outcompete organisms that generalize. This is true for cells, for hospitals, for hedge funds, and for DAVs. [I]
The question is never "how smart is your model?" The question is "how well do your specialized models coordinate?" [S]