RFC-SensOS21

Implementation Status

Reference Implementation: GemminAI/nvs-kernel

Validation: EXP-8000 Integration Validation — PASS

Quality Gates

• Ruff — PASS
• MyPy Strict — PASS
• Pytest — 256 passed
• Regression — None

Implemented: RuleLevel, Law, Policy, GovernanceEngine (nvs_kernel/governance/).

---

Abstract

This specification defines the logical structure of the Kernel Executive’s top-level Governance Layer, and the strict separation model between absolute constraints (Law) and adaptable operating policy (Policy). It resolves the opacity and ambiguity of “prompting policy” in conventional LLM agent systems by organizing rules of differing bindingness into a seven-tier hierarchy: Vision, Mission, Law, Policy, Goal, Directive, Instruction.

Three-Tier Layer Architecture

SensOS classifies every Kernel subsystem and entity into three responsibility layers:

1. Governance Layer — Vision, Mission, Law (External / Organizational / Kernel), Policy (Resource / Model Routing / Security / Cost), Goal (see RFC-SensOS23)
2. Execution Layer — Directive, Instruction, Workspace (RFC-SensOS22), Process (RFC-SensOS19), Scheduler
3. Runtime Layer — Memory (RFC-SensOS13), Cache (RFC-SensOS22), Observation (RFC-SensOS16), Knowledge/HEKB (RFC-SensOS17), Resource Manager

This “Runtime Layer” is a Kernel-manager classification axis, distinct from the R0–R7 Observation→Control pipeline defined by RFC-SensOS12–19. The two must not be conflated.

Seven-Tier Semantic Rule Hierarchy

Level	Rule Kind	Bindingness	Responsibility
L1	Vision	Permanent, top-level	The system’s reason for existing and long-term values
L2	Mission	Purpose / mid-to-long-term	Concrete mission a system or tenant must achieve
L3	Law	Inviolable	A rule that must never be broken by any exception or optimization
L4	Policy	Adaptable	Operating policy adjustable to budget, performance, environment
L5	Goal	Task-dependent, dynamic	Target to be achieved in a session/project (RFC-SensOS23)
L6	Directive	Execution-plan level	Abstract action plan an agent formulates to reach a Goal
L7	Instruction	Atomic-command level	Smallest execution step: tool call, prompt output, memory write

This L1–L7 scale is unrelated to the nvs-kernel Control Tier (Tier.L0_NONE…Tier.L7_HALT). The two share only overlapping digits, not meaning.

Law categories

• External Law — public regulation (GDPR, EU AI Act, copyright law, industry standards)
• Organizational Law — internal rules (security policy, contract terms, NDAs)
• Kernel Law — the SensOS constitution (Kernel Constitution, binding RFC clauses, physical/ethical safety guardrails)

Policy categories

Resource, Scheduling, Cache, Model Routing, Security, Cost.

Pre-Execution Validation Flow

Before an Instruction reaches physical execution (Runtime Layer), the governance validation engine runs a bottom-up constraint check:

Valid(I) = LawCheck(I) ∧ PolicyCheck(I) ∧ GoalCheck(I)

Order: Goal/Directive validation → Policy validation → Law validation (the final, absolute gate).

Violation handling

• Law Violation — unconditional Abort, regardless of Goal or Policy pressure. Recorded as a Kernel Fault; context is discarded.
• Policy Violation — may be exceptionally overridden by a privileged process (emergency debugging, manual override, priority exception).

Relationship to Other RFCs

• RFC-SensOS22 (Domain Separation) adopts this Governance/Execution/Runtime three-layer split and reorganizes Kernel manager responsibilities around it.
• RFC-SensOS23 (Goal Manager) tracks, decomposes, and evaluates this specification’s L5 Goal level, expanding it into Directives.

---

Validated by

EXP-8000 — Law Enforcement Primacy (EXP-8001): a CRITICAL-priority Goal whose plan violates a registered Law is unconditionally rejected (FAILED), independent of priority.