Editorial Process
Every SYLO article runs through the same five-stage pipeline before it reaches the public site. The pipeline is documented in detail in our internal editorial-policy.md, which lives in the SYLO content-pipeline repository and is maintained by the editorial team. This page is the public summary so readers, journalists, and search quality raters can understand how the content was produced without having to take the claim on faith.
1. Research
A single article starts from a single keyword, for example, “relationship stress triggers” or “vagus nerve breathing”. The research agent queries a curated set of sources: PubMed for clinical and neuroscience literature, Semantic Scholar and Google Scholar for broader academic coverage, and Exa for reputable web material when peer- reviewed coverage alone would leave gaps.
Every candidate source passes through a TF-IDF relevance gate that rejects anything that does not materially overlap with the article’s keyword. Surviving sources are bundled into a typed SourceBundle and classified into three tiers: Tier 1 peer-reviewed primary research, Tier 2 systematic reviews and meta-analyses, and Tier 3 reputable secondary material. If the bundle does not contain enough Tier 1 or Tier 2 material to justify the claims an article would need to make, the article does not get written. The bundle is sent back to the research agent for another pass.
2. Writing
A large language model drafts the article from the SourceBundle. Every factual sentence in the draft must carry an inline citation in the [Sn] format. The pipeline enforces this deterministically in a post-processing pass, not via prompt engineering alone. Each citation resolves to a specific source, and the article ends with a Quellen (German) or Sources (English) section listing the full references, so readers can open the original studies.
The writer is also asked to include differentiation elements (counter-examples, limitations of the cited studies, and specific effect sizes) so the resulting article is not a bland summary of well-known talking points but adds something to the available coverage on the web.
3. Automated review
Before a draft ever reaches a human, it is checked by the reviewer agent. The automated review does four things: it fact-checks each cited claim against the underlying source to make sure the summary actually reflects the paper; it evaluates the YMYL (Your Money or Your Life) severity of any health claims; it runs a regex-based compliance gate against the German Heilmittelwerbegesetz (HWG) § 3 forbidden patterns; and it audits every citation link to confirm the URL resolves.
Failing any of these gates causes the pipeline to reject the draft and re-enter writing with a structured feedback block. After three failed retries, the article is held and escalated to the editorial team for manual review rather than shipped.
4. Human editorial review
The three named editors (Max Reck, Noah Schömann, and Kyr Khoroshko) take the AI-reviewed draft, read it end-to-end, and make the final call on publication. This is where judgement calls about tone, clarity, and framing happen: an article that technically passes every automated gate can still feel prescriptive, sensationalist, or ambiguous, and the editor sends it back until it reads respectfully for a non-clinical audience.
The editor’s byline is not a claim of medical review. Every SYLO author bio reiterates explicitly that the author does not hold a medical license and that medical authority on SYLO articles derives from the peer-reviewed cited sources, not from the author. The editor’s job is editorial integrity (accurate reflection of the literature, honest framing, readable prose), not clinical sign-off.
As part of this stage the editor also runs the per-article fake-author-prevention checklist owned by the Seelenfreund Founder/CMO: confirming that the byline matches the editor who actually reviewed the article, that no invented credentials leak into the bio, and that no article body text implies medical review by a named person when no such review took place. This checklist is the human gate that protects against the most damaging YMYL failure mode, fabricated expertise, which code alone cannot catch.
5. Publishing
Once the editor approves, the article is serialised to Markdown with enriched frontmatter (author key, reviewer key, cluster, tags, FAQs, translations, datePublished, dateModified) and uploaded to S3. A content hash is computed so that regenerations that do not change the body do not move dateModified unnecessarily. The upload triggers a GitHub repository_dispatch event, the SYLO website rebuilds statically on Vercel, and the new URL goes live once the edge deploy finishes.
All SYLO blog pages are statically generated; there is no runtime data fetch on the rendered article. That means the HTML a search engine sees is the HTML the reader sees, and the E-E-A-T JSON-LD (BlogPosting, Organization, Author, and, for YMYL clusters, MedicalWebPage) is embedded in the server-rendered page, not injected later from the client.
Where to find more
For the honest policy on medical review specifically (and why SYLO does not currently claim an in-house medical reviewer), see Medical Review Process. To report an error or outdated citation, see Corrections.