Pre-Publish Quality Gate for AI-Assisted Content

A pre-publish gate that blocks a merge in about four seconds and tells the writer which claim failed. It does not check whether the text reads as machine-written. Detector scores are the least reliable input available: Originality.ai and GPTZero both flag human writing by non-native English speakers often enough to make them useless as a publish criterion, and on Arabic they are worse. I will not build a gate that rewrites your team's sentences until a detector calms down. What the gate checks: Claim binding. Every sentence carrying a number, a date, a regulation, a price or a superlative must name a source in the front matter. Unsourced claims fail the build. This is the check that changes behaviour, because it is the one a writer cannot satisfy by editing style. Numeric drift. Figures in the draft are compared against your source of truth: the fee table, the rate card, the API reference, the shipping matrix. A page saying 2.9 percent when your pricing page says 2.4 fails, whoever wrote it. Near-duplication. Every draft is embedded and compared against your own published corpus. Cosine similarity above 0.92 against a live page is a merge conflict rather than a new page, and the gate names the URL to update instead. Structured data. Schema is validated against Schema.org and Google's rich results requirements before publish, not after a crawl error. Author entities must resolve to a real person; a page whose author cannot be verified does not ship. Disclosure and provenance. Model, version, retrieved sources and named reviewer are written into the record. Where your market requires disclosure of AI assistance, the field exists and is populated rather than remembered. Source integrity. Cited URLs are fetched. A source returning 404 or redirecting to a homepage fails, because a dead citation is worse than no citation. Where it runs: As a check on the content repository in GitHub Actions, or as a CMS webhook holding the document in a blocked state. Failures post to Slack quoting the failing claim and the check that caught it. Overrides exist, require a named person and a written reason, and are logged. A gate nobody can override gets switched off within a month, so making overrides visible beats making them impossible. What is not included: No detector evasion, no humanising pass, no rewriting to move a perplexity score. No plagiarism scanning against third-party corpora and no copyright opinion. No editing service: the gate reports, your writers fix. I do not tune thresholds indefinitely; you get the thresholds, the reasoning behind each, and the ability to change them. Who this is not for: Teams whose real problem is that they publish claims nobody sourced, and who want a tool that makes this acceptable. The gate will fail almost everything on day one. That is the finding, and some buyers do not want it. Calibration: Thresholds are set against your last 200 published pages so you see the failure rate before it lands on your writers. That run is part of the engagement and is usually the uncomfortable part of it. Two clients have paused publishing for a week after reading theirs, which I consider the gate working before a line of it shipped.