Not Every Article Needs a Fact-Check Loop: A YMYL Classifier That Cuts Cost 5x

When we first built our article pipeline, every piece of content went through the same production line: LLM drafting, a live web fact-check pass, and human review on top. The result was safe — and painfully slow and expensive. Then we looked at what we were actually protecting, and realized we were paying full insurance on articles that carried almost no risk at all.

The problem: paying the same premium for every article

Our cost per article for the full fact-check loop — model-powered web verification plus a human pass — was about $0.05. That sounds cheap until you multiply it by a content calendar that publishes every week. Worse, the loop was the bottleneck: every draft waited for verification before it could ship.

But not all articles are equally dangerous. A post about a supplement's dosage that hallucinates a number can hurt someone. A "how to pair fonts" article with a made-up stat is embarrassing but harmless. We were spending identical effort on both.

The fix: a rule-based YMYL classifier

We built a YMYL (Your Money or Your Life) classifier that runs before the fact-check stage. It uses four rule groups — plain wordlists covering money, health, legal, and product safety parameters — to sort every article into one of two buckets:

• High-risk YMYL — articles touching money, health, law, or product safety data go through the complete verification loop, unchanged.
• Normal — everything else skips the heavy re-verification pass and ships with light checks.

The classification is explainable by design: a wordlist is something a human can read and argue with, unlike a black-box model. If a post mentions a drug name or a refund policy, the wordlist catches it, and the rule tells you exactly why.

What changed

Per-article fact-check cost dropped from about $0.05 to roughly $0.01 — a 5x cut — because the expensive loop now runs only where the risk actually lives. The money we saved went back into producing more articles, not into the P&L.

The rules we refuse to bend

• High-risk content never skips verification. The classifier only decides which articles are cheap to ship; it never decides that dangerous content is safe.
• Rules must stay readable. Wordlists, not a neural net's intuition, so every classification has an audit trail.
• Savings go to output. The point of cheaper verification is more content, not a cheaper report.

Reusable checklist: YMYL classification in 4 steps

1. Define the high-risk wordlist. Start with the four rule groups: money, health, legal, product safety parameters.
2. Classify by rules. Any match routes the article to the high-risk path.
3. Split the loop. High-risk gets the full verification pipeline; normal articles get lightweight checks.
4. Calibrate continuously. Whenever a risky article slips into the cheap path or vice versa, fix the wordlist and log the change.

If you run a content pipeline that treats every post like it can move money, our blog planning system is built around this exact trade-off — and it will happily keep the heavy checks where they matter.