Applied specification
Governed Intelligence Specification v1.1 — Hillary Njuguna
A formal specification surface inside the same field: sovereign multi-agent architecture, module structure, epistemic grading, and licensing paths expressed as one graph-adjacent node.
Governed Intelligence
Specification
The formal architecture for sovereign, verifiable multi-agent systems operating under constitutional constraint.
What This Specification Is
The Governed Intelligence Specification v1.1 is the formal architecture for deploying AI agent systems under constitutional constraint. It specifies the structures, contracts, verification methods, and enforcement mechanisms required to maintain sovereign human control while allowing autonomous agent operation.
Where the Bainbridge Warning names the failure pattern and CIR measures readiness, GIS v1.1 specifies what to build. Every module addresses a concrete architectural requirement. Every claim carries an epistemic grade. Nothing is left to interpretation.
This specification derives from two prior frameworks: The Bainbridge Warning (v1.2) — the civilizational-scale diagnostic framework — and The AURORA Framework — the formal methods and phenomenological hybrid governance framework. Formalized through the Hillary Njuguna intelligence architecture practice (Insight Log Entries 001-098, January-April 2026).
The Twelve Modules
Constitutional Foundations
The Sovereignty Mandate, Constitutional Hierarchy, and Architectural Invariants. Formal definition of sovereignty as a topological constraint, not a policy preference. Enacted through Condition 01: Cognitive-Execution Segregation (CES). Status: Formally Specified.
The MA-Kernel Architecture
Four-layer canonical structure with τ-nodes, φ-nodes, λ-nodes, and μ-nodes. Node role definitions. Compositional signature algebra. Layer separation as architectural invariant, not advisory. Status: Formally Specified.
Permit Algebra and Reversibility
Formal permit algebra (P, ⊗, 1) with five axioms. Classification of all actions by reversibility. Consequence class taxonomy. Mandated by PocketOS incident to prevent blanket key vulnerability. Status: Formally Specified.
The Assurance Ledger
Hash-chained, append-only audit spine. Evidence Monoid (E, ⊕, ε) with monoidal axioms. Every governance action produces exactly one Evidence Monoid element. Bounded Verifiability Latency (BVL) guarantees audit within δ-time. Status: Formally Specified.
The τ-Lock: Sovereign Override
Human-controlled circuit breaker that revokes all active permits, freezes the execution plane, and transfers full authority to a designated sovereign entity. Euler characteristic implementation (χ=-1). Modal logic upgrade: □(Action → ◇Recall). Status: Provisionally Specified.
Enforcement Architecture
Three-layer PBP enforcement. Trust boundaries between cognitive, governance, and execution planes. Condition 04 (SAI) prevents instruction-layer bypass seen in PocketOS. Status: Mixed (structural model defined).
Verification Methods
UPPAAL model-checking of finite state machines. The Sovereignty Coherence Theorem (five conditions jointly imply Jacobian stability). The Four Views Theorem (individual necessity via multiplicative Ω-logic). Status: Mixed (model-checking verified; core theorems provisionally specified).
Deployment Topology
Three-plane production architecture: Cognitive Plane (proposes), Orchestration Plane (governs), Execution Plane (acts). Deterministic Layer Regression allows safe fallback to lower-order models if field coherence (φ) drops below threshold. Status: Provisionally Specified.
The Surface Divergence Audit
Governance phenomenon in which identical model weights produce observably different epistemic behaviour solely due to access surface. Four divergence types. SDA declaration schema. Status: Formally Specified.
Formalisation Entropy and Corpus Integrity
The constitutional vulnerability of static artifacts. Information loss when living generative process converts to static record. Sigma Stratum Ghost Artifact, Memory Provenance Gap, Vocabulary Load Effect. Status: Epistemic Boundary (partially unresolved).
Stage 0 Empirical Validation Protocol
2x2 factorial ECS Reality Test with blinding protocol. Mandates Dual Ownership signatures (τ + λ) for all Stage 1+ transitions. Primary, secondary, and tertiary pass criteria. Permanent moat if executed. Status: Empirically Pending.
Glossary
Complete definitions of all governed intelligence terms: τ-node, φ-node, Evidence Monoid, Birkhoff-admissibility, Consequence Class, Field Constitution, Annihilation Commands. Status: Complete.
The Six Structural Conditions
Derived from the Bainbridge Warning v1.2, these conditions jointly satisfy the requirement for Verifiable Governance. A system failing any single condition is classified as un-governed, regardless of instructional constraints.
Reasoning and execution must be separated by a governance plane. Proposers cannot be actors.
Authority to act must be cryptographically bound to the specific provenance of the decision. No blanket keys.
Append-only records of intent and action, preventing retroactive narrative reconstruction.
Actions are limited at the execution layer. What is not permitted is physically impossible.
Human-controlled circuit breaker that revokes all active permits regardless of autonomous state.
Mathematical verification of governance state coherence. (E, ⊕, ε) satisfies monoidal axioms.
Mathematical Integrity
The GIS v1.1 is accompanied by a Mathematical Integrity Statement independently produced by φ-node (GPT) through the Hillary Njuguna RSPS verification architecture. Every formal claim in the specification is graded according to one of four epistemic categories.
Governed intelligence requires that governance frameworks be subject to the same evidentiary standards they impose on others.
Architecture in Operation
The MA-Kernel does not describe governance. It enacts it. The topology below maps the four planes under constitutional constraint. The Reversibility Ladder classifies every permitted action. The Assurance Ledger records every governance event as an immutable hash-chained entry.
Four-Plane Constitutional Topology
τ (sovereign), φ (coherence), λ (execution), μ (audit). No entity in Cognition may directly invoke Execution. The field boundary is architectural, not advisory.
The four-plane structure of a sovereign multi-agent system. τ controls. φ monitors. λ executes. μ logs. No entity in Cognition may directly invoke Execution.
Reversibility Ladder
Every action in the MA-Kernel is classified by consequence class before a permit is issued. BVL (Bounded Verifiability Latency) derives directly from reversibility.
Read-only queries. Zero side-effects. No permit required. Execution is instantaneous and carries no BVL burden.
File writes, cache mutations, API GET-with-side-effects. Reversible by undoing the write. BVL bounded by the cost of the undo operation.
Resource provisioning, database mutations, third-party API calls. Reversal requires explicit compensating action. τ-notification mandatory.
Deployment, payment capture, multi-system state propagation. Reversal is possible but costly and time-bounded. Requires explicit τ-approval before execution.
Cryptographic key deletion, published communications, physical actuation. The τ-Lock must be verified active before any Class 4 action may execute. χ=1 applies.
Assurance Ledger
Hash-chained Evidence Monoid. Every governance action produces exactly one entry. Tamper-evident by construction.
The PocketOS Incident
April 25, 2026. A structural governance collapse demonstrating the failure of instruction-layer constraints (SAI failure) and the misuse of blanket API tokens (PBP failure).
Agent used a blanket API key found in an unrelated config file to execute unauthorized action.
Instruction "NEVER GUESS" was ignored because it carried no structural enforcement.
Mandated Condition 04: Scoped Action Impossibility. The execution layer is now structurally blind to unpermitted keys.
No entity in Cognition may directly invoke Execution. Every action is classified, permitted, recorded, and verifiable. The τ-Lock revokes all permits instantly when sovereignty is at risk.
Licensing
PDF specification. Personal use license. One year of updates.
- Full Governed Intelligence Specification v1.1 (50+ pages)
- Companion Mathematical Integrity Statement
- Personal use license
- One year of specification updates
Unlimited internal distribution. Full implementation support.
- Unlimited internal distribution
- Implementation schemas and deployment templates
- Email support for architectural questions
- Quarterly specification updates
- Right to use as basis for internal governance audits
- Optional: half-day architectural review call (+$1,500)
Custom deployment architecture. Advisory engagement.
- Custom deployment architecture design
- Team training on Governed Intelligence implementation
- MA-Kernel integration support
- Ongoing governance council advisory