System Architecture
From the world, through an envelope, to a governed decision
SECS is a machine that only acts on packets of signals — not on people. Those packets are called envelopes. Every envelope is either admitted (the decision runs) or annihilated (nothing happens). There is no soft middle, no “try again later with half the data.”
For everyone: think of a sealed form that says what the room needs, not who you are. The form travels; your name stays at the door. The machine either accepts the form and acts, or rejects it and leaves the world unchanged.
For builders: a deterministic, identity-free substrate runs a governed collapse pipeline
C = ε∘ρ∘λ∘ν∘α.
Same admissible input, same doctrine ⇒ same path and signed proof. Replayable decisions, not best-effort filtering.
Above the fast path sits a Neurotrophic OS — it watches rates and faults, keeps the machine in a healthy operating band (homeostasis), repairs broken bridges, and learns within frozen constitutional bounds. External systems (clinic, warehouse, robot cell, agency roster) connect only through certified adaptors that create envelopes from the real world.
What travels — envelopes move information
People often land on this page because the diagrams look clean — and still wonder what is actually moving. The answer is the envelope.
Real-world source of the idea. Paper mail: a letter carries a message; the envelope is the boundary and the addressing. In computing, HTTP requests, message-bus events, and gRPC calls are the same pattern — metadata + payload crossing a trust boundary. SECS hardens that pattern: the payload may only contain signals the decision needs; identity keys are structurally excluded, not merely “discouraged by policy.”
How an envelope is created. Almost never inside the substrate itself. An adaptor sits next to the real system (robot controller, WMS, clinic device gateway, labour-hire roster). It reads sensors and business events, strips forbidden fields, packages allowlisted signals, and often signs a certificate that binds that vertical’s doctrine. A human may fill a UI; the machine still only sees the envelope.
Signals (travel)
- Cell id, cycle count, peak torque, over-temp flag
- Shift need: room, count, window — not the worker’s Medicare number
- Risk flags, fault codes, window start/end
- Whatever the sector doctrine allowlists
Identity (stays local)
- userId, email, session token, badge photo
- Names in free text, device fingerprints
- Anything that answers “who is this person?”
- Your org may still know responsibility outside SECS
Robotics soak example. A manufacturing cell runs under load for hours. Torque, temperature, cycle count, and e-stop state are continuous signals. Those are not “who is logged into the HMI.” The adaptor builds an envelope each window. If the packet is well-formed and doctrine-aligned, the substrate can emit a governed decision (continue soak / hold cell / require human override). If identity leaks or required signals are missing, the packet is annihilated — the robot does not get a half-approved command.
You can practice building envelopes on the
Sovereign Terminal (briefing included) or
The Game. Same gate as production adaptors on secs-sovereign.fly.dev.
What this achieves in the real world
Architecture is not the product. The product is predictable, auditable action under pressure.
| Outcome | What you get outside the lab |
| Fail-closed decisions | Bad or incomplete packets do not partially execute. The world either gets a full governed action or nothing. |
| Identity stays local | The machine acts on need and state, not on who walked through the door. Privacy and liability lines stay at the adaptor. |
| Replayable audit | Same envelope + same doctrine ⇒ same path and HMAC-signed proof. Regulators and operators can re-run the decision, not trust a log story. |
| Stable under load | Homeostasis keeps rate, latency, and error bands inside doctrine set points so a spike does not silently rewrite behaviour. |
| Safe adaptation | Neurotrophic layers may tune parameters and topology only inside frozen constitutional bounds — the system cannot author new goals. |
| One law, many verticals | Clinic, warehouse, robot cell, agency multi-site: same admission algebra; different allowlists and certificates per adaptor. |
In short: envelopes move the information the decision needs; the constitution decides whether that information may become action; extinction returns the substrate to purity so the next cycle does not inherit half-state.
System Layers
Every envelope’s signal travels one path, top to bottom. External sources hit a physics-only perimeter; the Go Atom kernel dispatches; the TypeScript runtime runs the collapse chain; governance and topology bound what may change; the neurotrophic slow path observes without blocking.
flowchart TB
EXT["External World"] --> PERIM["Perimeter / Connector"]
PERIM --> ATOM["Atom kernel — Go"]
ATOM --> PIPE["Collapse pipeline α ν λ ρ ε"]
PIPE --> GOV["Governance G₀…G₄ + topology"]
GOV --> NEURO["Neurotrophic slow path A–E"]
NEURO -.->|"bounded feedback"| PIPE
PIPE --> ADP["Adaptors STABLE / VOLATILE"]
ADP --> EXT
classDef layer fill:#141820,stroke:#00B2FF,color:#e0e2e8
class EXT,PERIM,ATOM,PIPE,GOV,NEURO,ADP layer
Combined stack — the same irreversible traversal used by the Sovereign runtime
flowchart TB
subgraph EXT["External World"]
RT["Real-time streams"]
SIM["Simulation / test"]
HUM["Human overrides"]
DOM["Domain systems"]
end
subgraph BOUNDARY["Ingress — physics only"]
PERIM["Perimeter Engine"]
end
RT --> PERIM
SIM --> PERIM
HUM --> PERIM
DOM --> PERIM
PERIM -->|"admissible sparks"| CONN["Connector / Atom"]
classDef ext fill:#161920,stroke:#00B2FF,color:#e0e2e8
classDef gate fill:#1a2418,stroke:#48e662,color:#dfffe5
class EXT ext
class BOUNDARY,PERIM gate
flowchart LR
subgraph INGRESS["Connector"]
CONN["Connector Atom"]
IDBAN["13 forbidden fields stripped"]
end
subgraph ATOM["Atom kernel — Go"]
BOOT["Load doctrine"]
DISP["Dispatch · 7.05 ns/op"]
end
subgraph RUNTIME["Substrate — TypeScript"]
SUB["Spark · Accept · Route · React · Extinguish · Emit"]
end
CONN --> IDBAN --> BOOT --> DISP --> SUB
classDef go fill:#1a1f14,stroke:#48e662,color:#e0e2e8
classDef ts fill:#141820,stroke:#00B2FF,color:#e0e2e8
class ATOM,BOOT,DISP go
class RUNTIME,SUB ts
Identity-free ingress — 13 constitutional fields never enter the substrate (spec)
Collapse Pipeline
Plain language: the envelope arrives (Spark). The gate checks whether it is allowed at all (α). If yes, the payload is made well-formed (ν), sent down exactly one path (λ), reacted on (ρ), then wiped clean so nothing sticky remains (ε). Finally a signed proof is emitted. If any constitutional check fails, the packet is vetoed — annihilated, not queued for a softer second try.
Operators: Spark through Emit — the same emit chain as Sovereign. Greek operators (ανλρε) name the irreversible transforms inside Accept–Extinguish. Each signal traverses once and exits as resolved output, governance proof, or structural veto.
flowchart LR
SPARK["Spark in"] --> ALPHA["α Admissibility"]
ALPHA -->|pass| NU["ν Validation"]
ALPHA -->|veto| V1["v₁…v₆ partition"]
NU --> LAM["λ Routing"]
LAM --> RHO["ρ Reaction"]
RHO --> EPS["ε Extinction"]
EPS --> EMIT["Emit proof"]
classDef pipe fill:#121820,stroke:#00B2FF,color:#e0e2e8
classDef veto fill:#2a1212,stroke:#FF5E00,color:#ffd8c8
class SPARK,ALPHA,NU,LAM,RHO,EPS,EMIT pipe
class V1 veto
Diagram above is the canonical path; hover each step for operator-level detail.
C² = C — collapse is a projection.
C¹⁻ ∄ — irreversible at every stage.
Vetoed signals are structurally impossible, not filtered or retried.
Exhaustive Veto Partition
Per the Collapse Algebra and Governance Veto Matrix: six classes partition every inadmissible input. No unclassified failures. Severity lattice v₅ ≻ v₁ ≻ v₄ ≻ v₃ ≻ v₂ ≻ v₆. Full definitions on the Glossary.
| v₁ | Axiom Violation — foundational law breach; collapse domain restriction |
| v₂ | Drift Introduction — would shift trajectory away from the fixed point φ̅ |
| v₃ | Identity Emergence — payload carries identity content (π ∩ ℕ ≠ ∅) |
| v₄ | Invariant Breach — would falsify a constitutional invariant |
| v₅ | Corruption Attempt — targets the constitutional layer directly |
| v₆ | Deployment Instability — would render deployment state inconsistent |
Governance Filtration
Constitutional layers. Each inherits constraints from below. Doctrine loaded at boot, immutable at runtime. No runtime self-modification of governance rules. G₀–G₄ here follows the SECS hierarchy (Principles → Algebra → SAC Axioms → Surface → Envelopes); the six veto classes above are the exhaustive partition at the admissibility gate, not a four-way summary.
flowchart TB
subgraph GOV["Governance filtration G₀ ⊆ G₁ ⊆ G₂ ⊆ G₃ ⊆ G₄"]
G0["G₀ Principles"] --> G1["G₁ Algebra"] --> G2["G₂ Axioms"] --> G3["G₃ Surface"] --> G4["G₄ Envelopes"]
end
subgraph SURFACES["Runtime surfaces"]
FROZEN["Constraint surface — FROZEN"]
MUTABLE["Adaptation surface — MUTABLE"]
end
subgraph TOPO["Topology seed graph"]
S1["Seed A"] <-->|"governed adjacency"| S2
S2 <-->|"governed adjacency"| S3["Seed C"]
end
G4 --> FROZEN
FROZEN -.->|"bounds"| MUTABLE
FROZEN --> TOPO
classDef gov fill:#1f1814,stroke:#FF5E00,color:#e0e2e8
classDef topo fill:#141820,stroke:#00B2FF,color:#e0e2e8
class GOV,G0,G1,G2,G3,G4,FROZEN gov
class MUTABLE,TOPO,S1,S2,S3 topo
Constraint Surface
- FROZEN — defines what is admissible
- Cannot be modified at runtime
- Loaded from doctrine at boot
Adaptation Surface
- MUTABLE within bounds
- Parameters the neurotrophic layer can modify
- Bounded by the constraint surface
Example: FROZEN — HOSTILE classifications require human override before release.
MUTABLE — anomalyThreshold may be tuned within adaptor bounds, not bypassed.
Neurotrophic Layer — Governed Adaptation Live Jun 2026
The neurotrophic layer sits above the fast path and observes its output. It never interrupts the fast path. It learns from it, adapts parameters within bounds, and feeds changes back — but the fast path continues running unchanged through every adaptation cycle. Phase E (meta-learning & integration) shipped and live-wired — see Observed adaptation. Browser compute: collapse propagation · compute demos.
flowchart TB
subgraph FAST["Fast path — collapse pipeline"]
FP["α → ν → λ → ρ → ε — unchanged during adaptation"]
end
subgraph SLOW["Slow path — neurotrophic layer"]
A["Phase A Homeostasis"] --> B["Phase B Plasticity"] --> C["Phase C Fault repair"]
C --> D["Phase D Temporal learning"] --> E["Phase E Meta-learning"]
end
FP -->|"observe metrics"| A
E -->|"bounded feedback"| MUTABLE["Adaptation surface"]
MUTABLE -.->|"never blocks"| FP
classDef fast fill:#121820,stroke:#00B2FF,color:#e0e2e8
classDef slow fill:#1a2418,stroke:#48e662,color:#e0e2e8
class FAST,FP fast
class SLOW,A,B,C,D,E,MUTABLE slow
Phase A — Governed Homeostasis
What it is: keeping the living system inside a healthy band — the same idea as body temperature or a plant’s coolant loop, applied to decision machinery.
What it does here: observes fast-path output (rate, latency, error, throughput), compares against doctrine set points, and reweights operational parameters within constitutional bounds. The slow path watches; the fast path never waits on learning.
Real-world win: a traffic spike, a noisy robot soak, or a flaky vertical does not silently rewrite how the gate behaves. You stay inside the band the constitution already approved.
Phase B — Structural Plasticity
What it does: grows and prunes topology nodes — adds seeds, strengthens useful connections, removes underperforming paths. All growth is bounded by the frozen constraint surface.
Real-world win: structure can evolve (new corridors in a warehouse graph, new soak routes) without the system inventing powers the constitution never granted.
Phase C — Cross-Domain Fault Repair
What it does: astrocyte-model mediator. Detects faults across domain boundaries, propagates bridge signals, repairs damaged circuits before learning proceeds. Learning on damaged circuits produces drift — Phase C closes that gap.
Real-world win: if an adaptor starts emitting malformed envelopes, Phase C can isolate the bridge and repair it before temporal learning treats bad traffic as truth.
Phase D — Temporal Learning
Learns from history. Applies reward mechanisms. Balances beneficial outcomes against hostile ones. Accumulates evidence until prediction confidence reaches threshold. Adjusts weights, not constitutional rules.
Real-world win: better prediction of which envelopes and paths work under load — without rewriting what is allowed.
Phase E — Meta-Learning & Integration
Learns how to learn. Adjusts learning parameters (not logic). Coordinates across all four preceding phases. Reward functions are Founder-defined only — the system cannot author its own goals. Topology-aware learning adapts based on structural context.
Real-world win: adaptation remains instrumented and bounded — useful under long soaks and multi-site agency load, never “the model decided its own purpose.”
External Verticals
Industry adaptors connect over HTTP only — the substrate never forks per vertical. STABLE/VOLATILE profiles, the adaptor diagram, compliance proofs, and per-vertical configuration live on Vertical Surfaces. Governed workflow walkthroughs are on Workflow Demos.
Observability — Dual-Lane Cockpit
Two independent cockpits. Same KPI set. Different data sources. Zero shared state. Stable lane is for governed production traffic; Volatile lane is for exploratory or high-variance environments where wider timing surfaces are intentional.
Stable Cockpit
Rate Surface
- Current RPS
- Min / Max / Avg
- Historical Trend
Timing Surface
- P50 / P95 / P99
- Jitter
Reliability Surface
- Error Rate / Count
- Success Rate
Governance Surface
- Throttle Rate
- Capacity Utilisation
- Accepted / Throttled
Muted · Precise · Minimal
Volatile Cockpit
Rate Surface
- Current RPS
- Min / Max / Avg
- Historical Trend
Timing Surface
- P50 / P95 / P99
- Jitter
Reliability Surface
- Error Rate / Count
- Success Rate
Governance Surface
- Throttle Rate
- Capacity Utilisation
- Accepted / Throttled
Expressive · Dynamic · Wide
Structural Properties
| Property | Enforcement |
|---|---|
| Deterministic | Same input → same output. Always. No unseeded randomness. |
| Identity-free | No personal identifiers in the substrate. Constitutional. |
| Irreversible | Collapse cannot be undone. C¹⁻ does not exist. |
| Idempotent | C² = C. Collapsing a collapsed state changes nothing. |
| Governed | All adaptation bounded by frozen constraint surface. |
| Additive | Each neurotrophic phase adds capability. Nothing modified. |
| Constitutional | Doctrine loaded at boot, immutable at runtime. |
End-to-End Path
Ingress through adaptation — one traversal. Collapse operators are detailed above; this is the full cycle including kernel dispatch and slow-path feedback.
The substrate never changes. Only the constraint surface changes.