AI agents are moving from answers to actions. Who holds the authority?
OpenKedge brings three questions into one architecture for governed AI agents. Hardknock tests what an agent learns in disposable environments. PCI examines what identity and authority may continue. PDDS and KedgeFlow govern when a proposal may change protected state with explicit, action-bound grants.
Research paradigms, open code, and reference implementations. Each workstream defines its own evidence and maturity; KedgeFlow v0.9 provides the open Rust reference gateway for bounded action.
Can this agent change this system, here and now?
- 01 · Proposal
An agent submits the exact action, parameters, target, and supporting evidence.
- 02 · Admission & grant
Admission checks policy and evidence; a separate issuer creates bounded permission for the approved action.
- 03 · Protected effect
The gateway claims authority once, validates the target guard, and tracks confirmed or unknown outcomes.
Hardknock · before deployment
What did controlled practice actually show?
PCI · across model change
What history and authority may continue?
You can rent intelligence. Authority must stay yours.
Models may be global. Execution authority must remain sovereign. A deployment must govern data disclosure, identity, policy, execution, and evidence under the institution's own obligations.
KedgeFlow
Programmable Control Plane
An action-bound authority protocol and Rust reference gateway, with executed MCP and etcd verification records.
Autonomous State Control Plane
Sovereign AI Systems
A deterministic control plane isolating unstable AI reasoning from live operations while preserving auditability.
From Sovereign Compute to Sovereign Execution
National AI Infrastructure
Maps sovereign execution across national infrastructure through federated institutional policy and localized authority.
Test what agents learn. Govern what they carry. Control what they change.
KedgeFlow defines the action control plane. PDDS studies its distributed-systems foundations, PCI governs identity continuity, and Hardknock supplies controlled experience for operator review.
What did it actually try?
Open-source, pre-alpha project
Hardknock gives agents disposable environments to test actions, compare alternatives, and record outcomes. It turns practice into evidence that people can inspect before relying on a lesson.
Give agents somewhere to fail before production.
What may it carry forward?
Persistent Cognitive Identity
PCI studies which memories, claims, relationships, and permissions a long-lived agent may retain across model and runtime changes. Acceptance and lineage must be reviewable.
What may it change?
Post-Deterministic Distributed Systems
PDDS studies how variable agent proposals become governed actions. KedgeFlow gives that boundary named components and interfaces for admission, grants, target enforcement, and outcome evidence.
These are complementary research and engineering workstreams, not a single deployed pipeline. Moving a Hardknock finding into PCI state or a PDDS decision requires explicit operator review and integration.
Read the claims. Check the boundaries.
Inspect KedgeFlow's executed verification, the PDDS and PCI research, and Hardknock's open code. Each result names the conditions and scope it actually covers.
KedgeFlow v0.9
The October 5 record reports 52 Rust tests passing with real three-member etcd and official MCP client checks. These establish local mechanics, with production deployment gates still open.
Post-Deterministic Distributed Systems
A systems model for coordinating AI agents, services, and people when proposals vary but consequential actions need clear authority.
Cognitive Admission Control
In 390 CAC trials, 120 effects completed with no modeled harmful effects observed. The result is specific to the evaluated setting.
Beyond Memory: Continuity Kernel
A model of governed state activation checked 2,808,230 reachable states with no invariant violations within that model.
Jev and the Agentic Control Plane
Jev separates semantic judgment from generation. We examine the systems boundary that comes next: separating probabilistic judgment from execution authority.
Open code · Pre-alpha
Hardknock makes the experience question testable.
Inspect the project and its implementation before treating a recorded lesson as reliable evidence.
An agent wants to shut down capacity. Who gets to say no?
Compare a direct tool call with an operator-defined admission path. This scripted example explains the control boundary; it is not a live deployment or a measured safety result.
Scripted example · synthetic records · no live cloud changes
Direct Execution to Cloud API
The AI agent holds ambient API credentials. Without pre-execution policy checks or admission control, mutations reach infrastructure directly.
{ "eventName": "TerminateInstances", "userAgent": "ai-agent-role/bot-session", "instancesSet": 25 nodes, "currentState": "shutting-down", }
Pre-Execution Admission & Fencing
The AI agent proposes candidate intent. OpenKedge Cognitive Admission Control (CAC) evaluates policy and environmental evidence before any authority is issued.
{ "intent": "shut_down_cloud_capacity", "activeInstances": 25 nodes, "policyCheck": "Scope exceeds operator limit", }
Give agents room to reason. Keep authority at the boundary.
Start with one consequential action and KedgeFlow's authority contract. Then map the PDDS evidence requirements, PCI continuity decisions, or Hardknock trials your workflow needs.