This document defines architectural primitives intended for use by standards bodies, procurement authorities, and governance frameworks. It is written as a reference artifact, not an opinion essay.
Purpose and Scope
This document defines a set of architectural primitives, modularity requirements, and conformance profiles intended to address a recurring failure mode in modern AI systems:
the delegation of execution authority to automated or agentic systems without enforceable, upstream constraints.
These definitions are architectural, not ethical.
They are designed to be implementation-agnostic, auditable, and suitable for procurement, governance, and security review contexts.
This document does not prescribe policy outcomes, model alignment techniques, or regulatory enforcement mechanisms. It provides a reference frame for identifying, constraining, and reasoning about execution authority in AI-enabled systems.
I. Architectural Primitives
The following primitives are non-negotiable building blocks for any system that claims to constrain agentic risk at the architectural level.
1. Admissibility Gate
An Admissibility Gate is a pre-execution control layer that determines whether a system is permitted to participate in a given domain, decision context, or workflow at all.
Characteristics:
evaluated before reasoning, planning, or tool invocation
domain-aware rather than confidence-based
capable of categorical refusal, not just escalation
An admissibility gate answers the question:
“Is this system allowed to act here?” — not “How confident is the output?”
2. Execution Boundary
An Execution Boundary is a hard architectural separation between:
reasoning, inference, or planning
and any action that alters external state
Reasoning components may generate suggestions, plans, or outputs, but must not directly invoke execution pathways.
Execution may occur only after:
explicit authorization
admissibility validation
or human confirmation, depending on system design
3. Temporal Authority Limit
A Temporal Authority Limit defines the maximum duration for which a system may retain execution authority without re-authorization.
This prevents:
indefinite agent persistence
authority drift over time
accumulation of unchecked contextual assumptions
Temporal limits apply regardless of model confidence or historical performance.
4. Context Integrity Requirement
A Context Integrity Requirement enforces safeguards against:
out-of-context interpretation
replay of partial or stale inputs
authority decisions based on fragmented state
Context integrity ensures that execution authority cannot be derived from:
truncated history
inferred intent
decontextualized signals
II. Modularity Requirements
Many real-world failures arise not from model behavior, but from tight coupling between system components. The following constraints are required to prevent authority leakage across layers.
Required Modularity Constraints
Models must not directly invoke execution APIs
Memory layers must not grant or infer authority
Tool access must be mediated by admissibility checks
Authorization logic must be external to the model
Agents must degrade to refusal, not escalation, when boundaries are reached
These constraints ensure that:
intelligence does not imply authority
persistence does not imply permission
confidence does not imply authorization
III. Conformance Profiles
Rather than imposing a single “standard,” this architecture defines conformance profiles that systems may voluntarily claim.
These profiles allow:
vendors to self-map
agencies to require minimum conformance
reviewers to assess risk exposure
DS-A: Advisory Only
No execution authority
Outputs are informational or recommendatory
No persistent state that enables action
Safe by default
DS-B: Human-Authorized Execution
Execution requires explicit human authorization per action
Admissibility gates enforced prior to authorization
Temporal authority limited to individual actions
DS-C: Autonomous Execution with Admissibility Controls
Execution permitted only within predefined domain boundaries
Hard admissibility gates enforced pre-execution
Continuous context integrity checks
Temporal authority limits enforced by design
Systems operating outside these profiles should be considered non-conforming for high-risk domains.
IV. Intended Use
This reference architecture is intended to be used as:
a background input to AI risk management frameworks
a framing memo for governance and standards discussions
a cited reference in procurement, oversight, and legal analysis
a design reference for builders of agentic systems
It is explicitly designed to complement, not replace, existing risk, ethics, and compliance frameworks.
V. Closing Note
The central claim of this document is narrow and technical:
AI risk is often introduced not by what systems predict or generate, but by where and when they are permitted to execute.
Architectural admissibility and execution constraints provide a control surface that downstream monitoring, auditing, and explanation cannot replace.
When Systems Wobble, It’s Rarely Random
AI hallucinations. Governance failures. Strategy drift.
Different symptoms — same architectural failure.
Over the past year, I’ve mapped a repeatable failure pattern across AI systems, institutions, markets, and organizations, formalized as the Drift Stack.
The diagnostic identifies which layer is failing — and why coherence is being lost.
Drift Architecture Diagnostic Assessment — $250
A focused 30-minute architectural review to determine whether the issue sits in:
Identity
Frame
Boundary
Drift
External Correction
If there’s a deeper structural issue, it becomes visible quickly.
If not, you leave with clarity.
👉 Full work index: https://www.samirac.com/start-reading
👉 Drift Assessment: https://www.samirac.com/drift-assessment
—
Chris Ciappa
Founder & Chief Architect, Samirac Partners LLC
Drift Stack™ · SAQ™ · dAIsy™ · Mind-Mesch™
