How does a topic become a published post?

The pipeline proposes topics for each site on one of three weekly tracks, and an editor approves the ones to write in the Apex dashboard. A write run every fifteen minutes picks up approved topics, has Claude write each post, adds the hero image, and runs the quality gate. A passing post is committed to the site's repository, or opened as a pull request on sites that review each post before it merges, and the site's deploy publishes it.

What does the quality gate check?

A validator blocks a draft with a missing hero image, too few internal links, a post link that doesn't match the site's URL format, a length outside the site's range, placeholder text, or an FAQ that repeats the answer capsule. A deterministic AEO score then has to meet the site's threshold. A draft below it is held for an editor instead of being rewritten, and a draft that fails the validator three times is retired.

How are hero images chosen?

Most sites get a hero from OpenAI's gpt-image-1, directed by a per-site art brief so the images stay on-brand. Sites whose imagery rules call for real photography get a Creative Commons photograph instead: licensed for reuse, never repeated on the same site, and credited in the post. Sites that produce their own heroes skip generation.

How does the pipeline avoid duplicate topics?

Before a topic reaches approval, the proposer checks it against every post the site has already published or queued. Topics rotate through the site's content clusters, and where the site runs an AI concierge, the questions its visitors asked are part of the brief.

Content Pipeline — AI Publishing, Model-Checked in TLA+

Q: What does the TLA+ model of the write path guarantee?

As modeled in TLA+, no write run outlives its time limit, at most one run holds the lock at a time, a cancelled job never runs, the worker never starts a job nobody is waiting for, and every attempt that ends without a published post counts toward the retry cap. The TLC model checker explores more than twelve million states for each property, and the run durations the model assumes are checked against thirty days of production logs. The checker must also reject a deliberately broken copy of the model for each property, and a faster profile of the check re-runs whenever the pipeline code changes.

Topic → approval → draft → quality gate → publish across about twenty sites, on a write path model-checked in TLA+.

Abstract

The Content Pipeline is the production system that writes and publishes articles for about twenty sites, each on one of three weekly schedules, across iSimplifyMe's own editorial network and client sites. It proposes topics, drafts each approved one in the site's voice, adds a hero image, and commits only what passes a structural validator and meets the site's AEO threshold.

On most sites, an editor's topic approval is the only human step. The write path that takes an approved topic to a committed post is model-checked in TLA+.

Problem

Producing content for many sites at once is operationally heavy, because every site has its own voice, calendar, and structural rules. Batch AI generation without editorial controls publishes pages that break those rules and don't build authority over time.

A pipeline that runs unattended also has to be right about timing. Scheduled runs, their retries, and other jobs share one worker queue, and a timing flaw surfaces only on the rare day their durations line up, a case that tests seldom reach.

Approach

Topics and approval

Each site's topic run fires on one of three weekly tracks — Monday and Thursday, Tuesday and Friday, or Wednesday — and proposes new topics within the site's content clusters. Every proposal is checked against what the site has already published or queued, and where the site runs an AI concierge, the questions its visitors asked inform the proposals.

An editor approves topics in the Apex dashboard. From there, the pipeline writes, gates, and commits unattended.

Writing and imagery

A write run starts every fifteen minutes, takes up to two approved topics, and hands each one to a scheduled worker that calls Claude with the site's voice, format, and linking rules. Each run holds a singleton lock and budgets its time, so no two runs overlap and every wait on the worker ends before the run's time limit.

Hero images come from OpenAI's gpt-image-1, directed by a per-site art brief. For a site whose imagery rules call for real photography, the pipeline picks a licensed Creative Commons photograph instead and writes its credit into the post.

Quality gate

Before anything is committed, each draft goes through a structural validator and a deterministic AEO score. The validator blocks any draft with a missing hero, too few internal links, a post link that doesn't match the site's URL format, a length outside the site's range, placeholder text, or an FAQ that repeats the answer capsule.

A draft that fails the validator three times is retired, and one that scores below its site's AEO threshold is held for an editor rather than rewritten. Some sites also require a passing score from an AI expert panel before commit, and some land each post as a pull request, so the site's own checks and reviewer sign off as well.

After a post is committed, the pipeline adds a link to it from up to three older posts on the same site whose text already names its topic. An hourly audit then checks each newly deployed page's structured data and answer capsule.

Verification

A formal model in TLA+, written from the code, covers the write path: the scheduler, the write run, its singleton lock, the job queue, and the worker. The TLC model checker explores more than twelve million states for each property, and the run durations the model assumes are checked against thirty days of production logs.

As modeled, the write path guarantees five properties:

Time budget — no write run outlives its time limit, so every run sends its summary and releases its lock
One run at a time — at most one run holds the lock at any moment
Cancelled stays cancelled — a job its requester cancelled never runs
No orphaned work — the worker never starts a job nobody is still waiting for
Counted retries — every attempt that ends without a published post counts toward the retry cap

Each property is also checked against a deliberately broken copy of the model, which the checker must reject; that shows the check can fail. A faster profile of the same check re-runs whenever the pipeline's code changes, and a test fails if a timing constant in the code drifts from the model.

Status

Writes for about twenty sites on three weekly tracks, with a write run every fifteen minutes (as of 2026-09-23).
The validator has gated posts since 2026-04-16, and the per-site AEO threshold since 2026-07-16.
Hero images: gpt-image-1 with per-site art direction, and credited Creative Commons photographs for sites that require real photography.
An hourly heartbeat alerts only when the pipeline degrades, and each weekday morning a digest reports the new topics by site.
The write path is model-checked in TLA+ (shipped 2026-09-23).

The same pipeline writes for client sites through Authority Engineering, iSimplifyMe's content infrastructure practice.

Content Pipeline

What is the iSimplifyMe Content Pipeline?