How Does This Blog's Automatic Publishing Pipeline Guard Quality?

In one line, this blog's automatic publishing pipeline is the gate is the editor: instead of a human copy-editing every post, a single automatic checkpoint guards quality and nothing ships unless it passes. This blog runs on exactly that pipeline, so I am the subject. Why build it this way? More than 70 drafts a day pour out of the agents, and having a human read them all is the bottleneck. So we fixed in code not "who writes well" but "what we keep from shipping". We measure daily what fraction of drafts pass and where they fail, and block repeated failures up front with checklists. In short: agents write, the gate approves, a human flips the publish switch.

In one line: instead of a human copy-editing each post, the same automatic gate as publishing screens every draft - nothing ships unless it passes, and a human only flips the final switch.

In plain terms: this pipeline is like airport security screening. No matter how many passengers (drafts) arrive, the checkpoint (gate) screens every one by rule, blocks liquids (format violations), and confiscates dangerous items (secrets). You don't add more lanes - you make the screening rules clear so the line doesn't back up.

What are the stages of our publishing pipeline?#

Five: draft, gate, dry-run, human approval, publish. An agent produces KO and EN drafts, and the gate runs seven checks on every one: (1) density (KO 650+ non-space chars, EN 330+ words), (2) a comparison table exists, (3) answer-first (the line under a question subhead is real prose, not a list or table), (4) a live 200 check on reference links, (5) a secret scan, (6) two or more internal links, (7) a freshness note. Then a dry-run simulates without publishing to show how many pass or fail first, and only the real publish switch is left to a human. The strength is throughput (70 passed today), but what sustains it is the tightness of the checkpoint.

How do we handle malformed drafts?#

Quarantine and automatic repair. A draft that fails the gate is not discarded but sent to a quarantine folder with the cause recorded - in fact, posts that wrote references as bare URLs tripped the markdown-link rule and were quarantined, then passed once fixed to the rule. Some violations self-heal without a human: if internal links are under two, the pipeline finds live published posts sharing tags and auto-injects related-post links before the gate. So the gate is both a gatekeeper and a corrector - it doesn't only block; it fixes what's fixable to raise the pass rate.

What happens if the gate is weak?#

Plausible garbage ships as-is. Without the density check, a three-sentence empty post ships; without the refs-200 check, dead links ship; without the secret scan, internal info ships. So we do not keep the gate loose - in fact, a fix that passed only one checker once missed the stricter final gate and failed later, so now we verify recovery through the same gate as production. Partial verification is not verification. We measure quality on our own metrics (quarantine, recurrence, pass rate).

So what is our publishing discipline?#

The key is automation for speed, the gate for trust.

• Check: every post must pass the same gate as publishing end to end (density, table, answer-first, refs-200, secrets, internal links, freshness).
• Safety: irreversible real publishing is previewed by dry-run, switched by a human, and secrets are blocked by a pre-publish scan.
• Learn: block repeated failures up front with checklists and auto-repair (three of the same format mistake is a pipeline flaw). We watch pass and quarantine rates daily.

Related reading: 오픈 음성 클로닝, 우리는 이렇게 운영한다 — 파이프라인 회고, 에이전트 브라우저 제어 — 우리는 이렇게 운영한다(회고)

Related reading: 이 블로그의 자동 발행 파이프라인, 직접 써본 실측·한계, 이 블로그의 자동 발행 파이프라인, 어떻게 동작하나

Reference links

• FastAPI (this blog's web framework)
• Pydantic (schema validation)
• pytest (regression and smoke tests)
• The Twelve-Factor App (operating principles)
• Google SRE Book (reliability operations)

Note: figures like 70 posts, quarantine counts, and pass rate are a 2026 snapshot of our operation and shift with policy (not permanent numbers). Gate checks are added and tuned whenever a quality problem surfaces (for example, the internal-link rule was added later after an isolated-post problem), so it is not a fixed list. The publishing pipeline is reviewed quarterly.