Status: pre-launch. This page describes target design; this document does not establish a deployed service.
Skyzai.ORG — Three Layers
Current reading — 2026-09-27
The sensor, inference and action allocation below is an interface design illustration, not proof that every adapter exists. Circle supplies sourced observations; forecasts preserve uncertainty. Signing, authority, execution and outcomes are separate guarantees. [I]
This reader boundary does not migrate interfaces, choose unresolved terms or prove deployment. See current status.
Layer 1 Perception. Layer 2 Cognition. Layer 3 Action. Every sensor, every model, every actuator. The constitutional firewall that prevents cognitive contamination.
The Architecture
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.
Three layers. Three interface contracts. Three independent failure domains. One organism. [S]
Each layer is agnostic to the implementation details of the layers above and below it. A webcam and a Bloomberg terminal both emit observations to Layer 1. A BitNet ternary model and a GPT-class transformer both process signals in Layer 2. A Hedera wallet and a robotic arm both execute commands from Layer 3. [S]
Layer 1: PERCEPTION — Sensor-Agnostic (Circle)
Circle does not care what generated the signal. It cares whether the signal is verified. [S]
Inbound Feeds
| Source Type | Examples | Modality |
|---|---|---|
| Social | Nostr relays, Twitter, Reddit | Text |
| Financial | On-chain events, bank feeds, Bloomberg | Numeric |
| Intelligence | OSINT scrapers, news APIs, government databases | Mixed |
| Physiological | Wearable biometrics, HRV, SpO2, temperature | Numeric |
| Visual | Camera feeds, webcams, satellites, LIDAR | Image |
| Auditory | Microphone arrays, voice streams | Audio |
| Environmental | IoT sensors, weather stations, smart homes | Numeric |
Every feed enters the same pipeline: [S]
Raw feed
→ Source validation (is this source legitimate?)
→ Claim extraction (what does this signal assert?)
→ Fact attestation (can this claim be independently verified?)
→ Evidence-tier classification (see [Evidence Tiers](14-evidence-tiers.md))
→ Typed, timestamped, evidence-tagged Observation record
Downstream consumers never know — and never need to know — whether the observation came from a satellite or a smartphone. [S]
Interface Contract
Observation {
source_id: string // Who/what produced this signal
timestamp: ISO8601 // When it was produced
claim: string // What the signal asserts
evidence_tier: string // [C]/[D]/[S]/[B]/[E] — how well the claim is supported
signature: bytes // Source authenticity proof
}
EigenTrust Scoring
Every source receives an EigenTrust reputation score that tracks the historical accuracy of its claims. High-trust sources propagate faster. Low-trust sources require more corroboration. The score is continuous and self-correcting — a source that begins producing false claims loses reputation automatically. [S]
Layer 2: COGNITION — Compute-Agnostic (Rosetta)
Rosetta is not a model. It is a translation layer. It takes typed observations from Layer 1 and routes them to whichever model architecture is best suited for the task. [S]
Supported Compute Backends
| Backend | L-Levels | Cost | Latency | Use Case |
|---|---|---|---|---|
| BitNet on-device (1B-7B) | L1, L3, L6 | $0 | 29ms | Always-on perception, filtering, compression |
| Haiku / 4o-mini | L1, L3, L6 | $ | ~1s | Cheap cloud perception and filtering |
| Sonnet / GPT-4-turbo | L3 | $$ | ~3s | Deductive precision, auditing |
| Opus / GPT-4o | L2, L4, L5, L7 | $$$ | ~5s | Exploration, decision, architecture |
| Gemini 1M | L5+ | $$$ | ~10s | Long-context diagnosis |
| Specialized domain models | L2-L5 | Varies | Varies | Medical, legal, financial, vision |
| SmallEBM / Boltzmann | SPECTRE routing | $0 | ms | Peer selection, energy landscapes |
| Prediction market aggregation | L2 | Varies | Varies | Crowd wisdom (not a model, but a compute output) |
The key property: no model has privileged access. A BitNet running on a $50 phone and Claude running on Anthropic's cluster both submit predictions to the same AXIOM convergence point. The prediction market does not know — and does not need to know — which model generated which prediction. Only the prediction's accuracy matters. [S]
Interface Contract
Prediction {
observation_ids: string[] // Which observations this addresses
probability: float // P(event) in [0, 1]
confidence_interval: [float, float] // Uncertainty bounds
model_class: string // Which model produced this
timestamp: ISO8601 // When it was produced
}
McGilchrist Routing
The Rosetta does not send every observation to every model. It routes using McGilchrist's hemispheric map: [S]
Signal arrives
→ LEFT lane: L1 on-device (cheap, fast, narrow)
→ Novelty detected?
→ NO: L3 filters, L4 decides. Done. (95% of cases)
→ YES: RIGHT lane: L2 explores (capable, broad, high-context)
→ L3 filters explorations
→ L4 decides
→ Blocked? → L5/L6/L7 evolutionary-loop escalation → L4*
The model doesn't know it's part of Skyzai. It just answers the prompt. The Rosetta system prompt is the coordination protocol. [S]
Layer 3: ACTION — Actuator-Agnostic (Skyzai)
Skyzai does not care what executes the command. It cares whether the execution was signed. [S]
Outbound Effectors
| Effector Type | Examples | Authorization |
|---|---|---|
| Hedera wallet | Token transfers, smart contract calls | Natural principal for a private act, or valid PRISM decision receipt for a public DAV |
| External wallets | WalletConnect (EVM, Solana, etc.) | Natural principal for a private act, or valid PRISM decision receipt for a public DAV |
| Robotic actuators | POS inventory, manufacturing equipment | Deterministic adapter consumes a typed authority receipt |
| API calls | Exchange orders, SaaS integrations | Applicable external mandate and exact consequence-bound authorization |
| RELAY broadcasts | Content distribution, alert propagation | Policy and non-authorizing PASS constrain proposals; neither authorizes execution |
| Smart home / IoT | Helios mesh commands | Applicable external mandate and exact consequence-bound authorization |
Execution consumes the applicable exact-bound authorization and produces an execution receipt. Missing receipts or timeouts require rail-specific investigation and recovery; they establish neither successful execution nor universal automatic reversal. Independent outcome evidence remains separate.
Interface Contract
Action {
instruction: string // What to do
authorization_chain: Auth[] // Who approved and why
k0_receipt_anchor: bytes // Where the receipt will anchor
effector_id: string // Who executes this
authority_receipt: bytes // typed natural-principal or PRISM authorization
}
Typed Authority at Every Level
| Context | Who Signs | What Constrains |
|---|---|---|
| Private DAV | Under applicable law, the Natural Person authorizes their own exact typed private act | eta = 0, Three Gates, constitutional rules |
| Public DAV | the signer(s), signer classes, threshold, and procedure required by applicable law, the entity's governing instrument, or a valid competent mandate bind the exact consequence in a complete valid bound PRISM decision receipt | VMOSK-A and policy constrain unsigned proposals but never authorize execution; PRISM records and verifies only; deterministic execution is limited to that exact consequence |
| Robot / Actuator | No signer; accepts only an exact receipt-bound instruction | Constitutional limits on autonomy |
| Agent wallet | No signer; stages unsigned proposals and accepts only an exact receipt-bound instruction | FlowWallet streams, P-score thresholds |
Where AI deliberation is used, it stages unsigned proposals; it is not a universal prerequisite for every action. Applicable mechanical constraints and exact-bound external authorization remain separate requirements. A Natural Person authorizes their own typed private act; a Legal Person acts through lawful external authority and required executors; and the signer(s), signer classes, threshold, and procedure 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; deterministic execution is limited to that exact consequence. The repository supplies no default signer count, signer class, or singleton fallback; the controlling signer or signers remain distinct from software and caste seats. [S]
Why Three Layers — The Constitutional Firewall
The three-layer separation enforces the pipeline at the infrastructure level: [S]
| Layer | Stage | Function | Failure Mode If Merged |
|---|---|---|---|
| Sensors | IS | Perception | Observations contaminated by desired outcomes |
| Compute | COULD | Prediction | Predictions anchored to sensor artifacts |
| Actuators | ACT | Execution | Actions bypass deliberation |
If sensors and compute share a pipeline, the system hallucinates — it "perceives" what its model expects. If compute and actuators share a pipeline, the system acts impulsively — it executes predictions before they are adjudicated. The three-layer boundary is a constitutional firewall. [S]
The Vertical — One Complete Cycle
1. Webcam captures customer at POS terminal (Layer 1: sensor)
2. On-device BitNet identifies product and prices it (Layer 2: compute)
3. Customer's wallet signs payment; POS receives (Layer 3: actuator)
4. Receipt posted to Hedera — K0: no receipt, no reality
Three layers. Three interfaces. Three independent failure domains. One organism. [S]
Agnosticism as Moat
| Scenario | Non-Agnostic System | Skyzai |
|---|---|---|
| Nostr relay goes offline | Circle breaks | Routes around it |
| OpenAI changes pricing | Rosetta breaks | Shifts to cheaper backend |
| Exchange API changes | Skyzai breaks | Routes through alternative effector |
| New sensor type ships | Requires rebuild | Plug in, same pipeline |
| New model ships | Requires retraining | Wrap with Rosetta, done |
Every sensor manufacturer, model provider, and wallet protocol becomes a plug-in rather than a dependency. The organism evolves its sensory apparatus, cognitive capacity, and motor capabilities independently. [S]
Every sensor that feeds it makes it see better. Every model that connects makes it think better. Every wallet that joins makes it act better.
Historical predecessor wording — retained on 2026-09-27, not active rules
The displaced passages are preserved verbatim for lineage. This retention date is not their original authorship date.
Predecessor passage 1
| API calls | Exchange orders, SaaS integrations | Ratified policy receipt plus typed executor | | RELAY broadcasts | Content distribution, alert propagation | Constitutional pass | | Smart home / IoT | Helios mesh commands | Natural principal or ratified policy receipt |
Predecessor passage 2
Every effector receives the same instruction format: a signed transaction envelope containing the action, the authorization chain (who approved this and why), and the K0 receipt anchor. The effector executes and returns a receipt. If no receipt returns within the timeout window, the action is reverted. [S]
Historical predecessor wording — retained on 2026-09-28, not active rules
Editorial reconciliation of this public source only; no upstream contract is amended.
Predecessor passage 1
All three substrates must agree for any action to execute: mechanical invariants, AI and caste deliberation limited to unsigned proposals, and typed authority.
Predecessor passage 2
limited to that exact consequence. The public DAV has no founder, chairman, owner, signatory, or single-human fallback. [S]