A quality gate that can say no
Volume tools publish whatever the model produces. Draftwave scores every draft against a rubric with hard floors — and rejects the ones that don't clear it. Nothing generic, thin, or risky ever reaches your site.
Stage 1 · Deterministic checks
- One H1, ≥ 3 H2s, no empty sections
- Length within brief ±25%, ≥ 600 words
- Meta title ≤ 60, description ≤ 155
- JSON-LD validates (Article + FAQPage)
- Keyword in H1 + first 100 words, density < 2%
- ≥ 2 internal links resolve, external links 2xx
- Dedup vs existing articles < 0.92 cosine
Stage 2 · LLM-as-judge
- Unique value present floor 3
- 4 / 5
- Accuracy risk floor 3
- 4 / 5
- AEO structure floor 3
- 5 / 5
- Depth · Originality · Readability avg ≥ 4.0
- avg 4.2
A failing draft doesn't reach this screen — it's rejected or auto-revised first.
Deterministic checks catch the cheap failures first
Before we spend a single judge token, fast code rules run for free: structure, length, meta limits, valid schema, keyword placement without stuffing, resolving links, and a de-duplication check against your existing articles. Anything that fails here is caught instantly — so the expensive evaluation only runs on drafts worth evaluating.
Hard floors override the average
A draft can average 4.5 out of 5 and still be rejected if unique value present scores a 1. The same floor applies to accuracy risk and AEO structure. This is what stops a polished-but-generic article — the exact kind Google's spam updates punish — from ever publishing. The benefit isn't a number; it's that nothing embarrassing lands under your name.
You stay in control
Passing the gate means pending_review,
not published. Articles land in a review queue with their scorecard attached — the unique
angle they added, the sources they cited, the AEO structure they used. You read, tweak,
and approve. Autopublish exists, but it's opt-in per site and rate-limited.
Frequently asked questions
Can the quality gate actually reject a draft?
Yes — that's the point. Drafts that fail the deterministic checks or fall below the hard floors on unique value, accuracy risk, or AEO structure are rejected or sent for a human to fix. They never publish automatically.
What is the 'unique value present' hard floor?
It's a non-negotiable minimum: a draft with no first-hand data, opinion, or example beyond restating the SERP fails regardless of how high its other scores are. This is the anti-generic rule that keeps thin content off your site.
Does the LLM decide whether to publish?
No. The model advises with a 1–5 score per criterion; code makes the pass/reject decision using fixed thresholds and hard floors. We also run the judge on a different model family than the drafter to reduce self-preference bias.
What happens to a draft that just misses?
If it clears the hard floors but scores in a middle band, the gate runs one bounded auto-retry of the edit step with the judge's suggested fixes. If it still doesn't pass, it goes to your queue as needs_attention.
Publish only what clears the gate.
Start free with 3 quality-gated articles — no credit card.