How variis uses AI

The short version

AI drafts. Humans review. Every claim is sourced. The AI never publishes on its own; a named editor signs off before anything appears on variis. If a sentence cannot be traced to a cited source, it does not run.

Which AI models we use

variis runs on two models from Anthropic.

Anthropic Claude Sonnet 5 handles the working load: research synthesis, first-pass drafting, and Gates 1 to 3 of our editorial pipeline (triage, drafting, and source verification). It is fast enough to sweep the day's signals across trials registries, preprint servers, and verified press feeds, and disciplined enough to draft news items that stay tethered to the supplied sources.

Anthropic Claude Opus 4.8 handles the harder judgement calls: verification of strategic claims, cross-checking numerical figures, and a second pass on Gate 3 when a story touches clinical outcomes, regulatory status, or company financials. Opus is slower and more expensive, so we reserve it for the claims where getting it wrong would cost a reader something.

We use both models through Anthropic's API. We do not train models on reader data, and reader analytics are never sent to the models.

What the AI does

• Ingests structured signals from a fixed list of named public sources, including ClinicalTrials.gov, Europe PMC, ANZCTR, HSA and TGA public registers, and a curated set of verified press feeds.
• Clusters related signals so that a single trial update, its preprint, and its press coverage arrive at the editor's desk as one story, not three.
• Drafts factual news items grounded strictly in the supplied sources. The model is instructed to refuse rather than invent when a source is missing.
• Produces a per-claim source-grounding record: every sentence in a draft is mapped to the specific source and passage it came from.
• Runs automated brand and voice checks against variis house style, including the banned-words list, sentence-case rules, and the evidence-grading conventions in our editorial standards.

What the AI does NOT do

• It does not set editorial policy. What variis covers, and how, is decided by the editorial team.
• It does not select which stories run without human input. Triage produces a shortlist; a human picks from it.
• It does not publish autonomously. Every article is reviewed and signed off by [Reviewer Name] before it appears on-site.
• It does not fabricate sources. If the grounding record is incomplete, the draft is rejected at Gate 3 and returned for rework or spiked.
• It does not make medical recommendations. variis reports on evidence; it does not prescribe.

The four gates

Every story passes through four gates before a human editor sees it.

1. Triage. Is this a real, novel longevity story with a clear Singapore or Australia angle?
2. Draft. Grounded generation from the clustered sources, with the per-claim record built as the draft is written.
3. Verify. Every claim is traced back to a cited source. Anything that cannot be traced is cut.
4. Brand. Voice, sentence case, evidence grading, and YMYL compliance are checked against house style.

Human review sits after the four gates, not inside them. The full model is documented at /editorial-standards.

When we get it wrong

Corrections are logged publicly at /corrections, with the original wording, the correction, and the date. If you spot an error in a variis article, write to corrections@variis.co and we will act on it.

Why we disclose this

Credibility in health and science reporting depends on transparency about method. Readers making decisions about their own longevity deserve to know what tools produced the words they are reading, where the claims came from, and who signed off. This page is how we answer those questions in public, and we will update it whenever the pipeline changes.