Research index
R4Research notev0.1
Human-Reviewed Agentic Content Pipeline
A preprint for planner/writer/worker orchestration with approval gates.
Shiva is useful only if agents accelerate drafts while humans retain publication control. This note scopes that boundary.
ShivaPublished 2026-07-03Updated 2026-07-03
Abstract
This research note defines the safe shape of Shiva: agent roles can research, outline, draft, and package content, but publication requires human review. The work is connected to the existing VNG blog/admin pipeline and Hermes draft endpoint. The claim is not autonomous publishing; it is supervised throughput with clear review states.
Claims
- Agentic content systems need review states, not hidden auto-publish paths.
- Research, drafting, and QA should be separate roles with evidence attached to each output.
- The safest first product is a draft accelerator for owned channels.
Methods
- Review of the current blog admin pipeline and Hermes draft endpoint.
- Scope split between planning, drafting, QA, and human approval.
- Content safety review against dark-pattern and low-evidence publishing risks.
Reproduce This
- Create a draft through the protected Hermes endpoint.
- Verify the draft lands in admin review state before publication.
- Confirm publication still requires a human admin action.
Limitations
- External source citation quality depends on the research provider configured for a run.
- The current public artifact is a research note, not an open agent demo.