An executable reference architecture for traceable, human-in-the-loop agentic publishing workflows.
This repository demonstrates how approved source boundaries, editorial memory, specialist agent stages, evaluation notes, and explicit human publication authority can be represented in code. It is maintained by Alomat LLC, a managed agentic publishing operations platform.
This is a public reference implementation—not a customer system, private Alomat source code, or a performance benchmark. It contains no customer data or private operational details.
Many publishing demos collapse research, writing, verification, and publication into one opaque prompt. This project makes the operating boundaries visible:
flowchart LR
A[Approved signal] --> B[Source verification]
B --> C[Editorial context]
C --> D[Content package]
D --> E{Human review}
E -->|Approve| F[Delivery]
E -->|Edit / reject / escalate| G[Stop or revise]
The central rule is simple: agents prepare; editors decide.
- Only configured sources enter the workflow.
- Primary references stay attached to the package.
- Editorial context is applied before content generation.
- Evaluation notes travel with the review package.
- Agent roles cannot approve or reject content.
- Delivery cannot run before explicit editor approval.
- Every accepted state transition creates an append-only event.
Requires Python 3.9 or newer and no third-party dependencies.
python3 -m unittest discover -s tests -v
python3 -m examples.managed_workflowpublishing_reference/workflow.py— state machine and invariantstests/test_workflow.py— executable control-boundary testsexamples/managed_workflow.py— end-to-end approved pathdocs/architecture.md— production architecture extension pointsdocs/evaluation-checklist.md— review and evaluation questionsdocs/threat-model.md— safety and abuse boundaries
MIT. See LICENSE.