Generation is one stage, not the product
Reliable content operations require research, retrieval, planning, evaluation, review, publication, feedback, and recovery around model output.
The system coordinates asynchronous content work across multi-provider LLM execution, PostgreSQL/pgvector, Redis/Celery, editorial state, WordPress delivery, and operational monitoring. Publication is fenced to the exact approved article revision so stale or revoked content cannot silently inherit prior approval.
Client / company work — ownership of company IP is not claimed
Company identity may remain withheld; displayed data is synthetic/sanitized; production/GA claims stay evidence-bound.
Publishing authority belongs to an approved revision, not to whatever content happens to be current.
Question · How do we let AI move content quickly without letting stale or unapproved state cross the publishing boundary?
APPLIES TO
Content Teams · Marketing Operations · SEO Teams · Publishers · High-volume structured content organizations
PROOF
Revision-bound publishing · RAG + multi-provider execution · controlled canary and rollback
CAPABILITIES
13 declared in the project record
Publishing authority follows the approved revision.
REVISION LINEAGE
Approval belongs to a revision, not to a topic forever.
audience · brand · topic intelligence
approved knowledge + first-party data
intent · outline · evidence · keywords
multi-provider workflow
quality + deterministic SEO
human release authority
idempotent adapters
cost · jobs · outcomes
SEO lift, traffic lift, conversion lift, and throughput savings are not claimed.
The constraints that matter.
Reliable content operations require research, retrieval, planning, evaluation, review, publication, feedback, and recovery around model output.
A newer edit must not silently inherit the release authority granted to an older article state.
Queued work can become stale after approval; the worker must fence revoked or superseded revisions before the external write.
Search Console and SEO signals can inform the next iteration without autonomously rewriting or promoting content.
What would make the preferred answer fail?
EVALUATION METHOD
A content workflow is reliable only when the exact approved revision—and no stale successor—can cross the publishing boundary.
Revision-bound review fixtures, stale/revoked approval tests, durable queue execution, provider-failure injection, publish verification, backup/restore checks, and canary/rollback gates.
What the evidence actually supports.
63 passed
Recorded repository validation; not a current production-traffic result.
7 passed
Focused unit validation; not a substitute for full supported launch acceptance.
PASS
Environment-sensitive warnings were tracked separately in the source reports.
10 Alembic revisions · single head 20260801_001
Schema topology evidence, not live migration proof against production data.
uses_llm=false · performs_network_requests=false · rewrites_content=false · publishes_content=false
Deterministic advisory layer; it does not autonomously modify or publish content.
Production-ready and deployed
SEO lift, traffic lift, conversion lift, and throughput savings are not claimed.