The Gantral Research Series
A formal body of work defining execution-time authority as infrastructure for agentic AI systems.
The series progresses from problem definition, to normative invariants, to reference implementation.
[ Execution Gap ] → [ Invariants ] → [ Implementation ]
The AI Execution Control Plane
Restoring Human Authority, Determinism, and Verifiability in AI-Driven Systems
This position paper defines the structural gap in execution-time governance. It argues that scalable AI adoption requires a distinct infrastructure layer responsible for pausing, enforcing, and replaying authority decisions.
Focus areas:
Execution-time governance
Deterministic replay
Separation of authority from intelligence
Control-plane framing
Audience:
Architects, governance leaders, regulators, and platform engineers.
Admissible Execution: Invariants for AI Execution Authority
This normative paper defines the minimum admissibility bar for execution-time authority.
It establishes non-negotiable invariants that must hold for authority decisions to survive adversarial review.
Defines:
Authority binding to execution instances
Replay independence
Immutable execution history
Enforcement over assertion
Separation of policy from authority
This document is vendor-neutral and implementation-independent.
Audience:
Auditors, risk leaders, standards bodies, governance teams.
Gantral: Implementation of an Admissible AI Execution Control Plane
Operational unification, deterministic authority, and verifiable chain-of-custody for agentic AI systems.
This paper operationalizes the admissibility invariants through a deterministic control plane architecture.
Key components:
Canonical authority state machine
Cryptographically chained commitment artifacts
Workflow and policy version binding
Deterministic replay independent of logs
Separation of policy from workflow implementation
Audience:
Enterprise platform teams, infrastructure architects, auditors.
How the Series Connects
Paper 1 defines the structural problem: execution-time governance is missing as infrastructure.
Paper 2 defines the admissibility bar: the invariants required for authority to survive scrutiny.
Paper 3 demonstrates enforcement: a concrete control plane implementation satisfying those invariants.
Together, these works define execution authority as a first-class infrastructure layer for AI systems.