Methodology

This page explains how we decide what to publish about each herpes vaccine candidate, where our information comes from, and how we keep it current. The short version: every status claim is traceable to a named source and carries a visible date.

What we track, and what we don't

We track the development status of herpes-simplex-virus (HSV) vaccine candidates — what stage each is at, who is developing it, and what the evidence does and doesn't show. We do not give medical advice, predict approval dates, rank candidates by promise, or recommend any product.

Vaccines are not the only line of HSV research. Antiviral drugs, antibody treatments and gene editing are aimed at virus already in the body rather than at training the immune system in advance, so they fall outside what this tracker covers and we never open a candidate page for one. Where that research is likely to be confused with a vaccine, we explain it as background instead: see gene editing and herpes, and why it isn't a vaccine.

Our sourcing hierarchy

For any factual claim we use the highest-quality source available, and we cite the tier we used:

1. ClinicalTrials.gov — the registry of record for whether a trial exists, its phase, and its recruitment status. This is the backbone of every status claim, cited by its NCT identifier.
2. Peer-reviewed literature (PubMed / Europe PMC) — for the underlying science and any published results.
3. Official primary announcements — sponsor press releases, regulatory (FDA/EMA) communications, and securities filings, for developments not yet in a registry or journal.
4. Reputable science journalism — for context and framing only, never as the sole source for a factual status claim.

We do not source from forums, supplement or "natural cure" sites, affiliate pages, social media, or any site selling a treatment. If a claim appears only in those places, we don't report it.

How we label phase and status

A candidate's phase uses the standard clinical-trial ladder (preclinical, phase 1, phase 1/2, phase 2, phase 3, approved), plus discontinued for programs that have stopped and unknown when the record is unclear. We take phase from the trial registry wherever possible. When a field can't be verified, we mark it unknown rather than guess.

Distinctions we preserve

Collapsing these is the most common way this topic gets reported wrong, so we keep them explicit on every candidate:

• HSV-1 vs HSV-2 — different epidemiology, often different candidates.
• Prophylactic vs therapeutic — preventing infection is a different goal from reducing outbreaks in someone already infected; a candidate can do one and not the other.
• Preclinical vs in-human — animal or laboratory results are not clinical results.
• A trial existing vs a trial succeeding — a phase 1 start is not evidence of efficacy.

Language we avoid

We don't use "cure," "breakthrough," "miracle," "game-changer," or any phrase promising a launch date. We state what the evidence shows and stop. Where something is unknown, we say so.

Keeping status current

Trial status drifts. We re-query ClinicalTrials.gov for the candidates we track and compare the registry's current phase and status against what's on the page, and we search the registry and the biomedical literature (via Europe PMC) for new HSV vaccine trials and papers. Any difference is reviewed before the page changes, by our Managing Agent (an AI) and afterward by the site editor (a person) — no script silently rewrites a health claim — and the page's "last updated" date moves only when the content actually changes. Registry facts are shown as a dated snapshot, kept separate from our editorial status.

How we use AI assistance

We use AI tools to help find new trials and papers, to draft neutral summaries, and to check each drafted statement against its source. We use them within strict limits:

• Every drafted statement is tied to the source's own words. Software confirms that each supporting passage appears word for word in the paper or registry record, and that every number we state appears in one of those passages. Anything that can't be matched is rejected.
• A second, separate AI check — run on its own and shown only the source and the drafted statements — looks for errors: overstated findings, animal results presented as human ones, the wrong herpes type.
• The Managing Agent approves; a person reviews afterward. No script publishes on its own. Our Managing Agent, an AI, reads each draft, its sources and both checks, and approves, edits or rejects it. The site editor reviews what was published afterward and can correct or reverse it. Before 26 October 2026 the editor approved each entry first.
• We label what kind of evidence it is. Each paper is marked as peer-reviewed, a preprint or a conference abstract, and by study type (laboratory, animal or human trial). Where our summary rests on an abstract alone, the page says so.

This is an editorial check, not peer review. Entries prepared this way carry a "How this entry was checked" note listing each statement and where in the source it comes from. The Evidence Desk page describes how new trials and papers are screened before the Managing Agent approves them.

How we review

Our monthly State of the evidence updates, and the weekly checks behind them, are done by a small team of AI agents, with a person reviewing afterward. These are AI agents, not people. Here is how the work is divided:

1. Specialist agents gather facts. One checks the trial registry, one reads new studies, and one looks at company and regulator announcements. They report what they found and where they found it. They do not say how important it is.
2. The Managing Agent reviews their findings. This AI editor is the only agent that judges how significant a finding is, and it uses fixed written criteria to do so. It can reject a specialist's finding, for example when the source is too weak or the claim goes further than the source does. A specialist cannot overrule it. It runs on a more capable AI model than the specialists, and it writes the monthly State of the evidence issue itself.
3. The Managing Agent approves and publishes; the site editor reviews afterward. When a draft passes the software checks, has a source for every claim and has no unresolved problem from a fresh-context review, the Managing Agent publishes it. It does not publish when a check fails, a claim has no source, or the change touches a discredited program, the site's scope or these review rules; those wait for the editor. The site editor reads published changes afterward and can correct or reverse them. Before 26 October 2026 the editor approved each change before it was published. The editor is not a physician. See About.

The agents follow the same rules as everything else on this site: the sourcing hierarchy above, the same distinctions, and the same banned words. Every statement in an update must trace to a named source. We do not rank or score vaccine candidates, and an update says what is known and what is not, not where a program is headed. This is an editorial check. It is not peer review. The software checks used on research entries are described in How we use AI assistance.

Frontier Analysis

The Frontier Analysis page is different from everything above, and is labeled that way. The Frontier AI, a top-tier model, looks across the whole record on this site for patterns, and proposes questions the record raises. The Critic AI, a different model, tries to find fault with each item and its notes are published with it. Software recomputes every number and checks every citation. An item the Critic AI finds unsupported is left off.

• Only what is already on this site counts as evidence. The agents use our own pages and the registry snapshot, not the open web.
• Questions are labeled as guesses. A "Hypothesis by Frontier AI" is not a finding, not a forecast and not advice. It never ranks vaccines and never says where a program is headed. Hypotheses are not cited anywhere else on the site.
• Small numbers are called small. Fewer than five independent programs is shown as cases, not as a pattern.
• The Managing Agent approves it; a person reviews afterward. Our Managing Agent, an AI, approves and publishes each issue, and the editor reads it afterward. It is an editorial check, not peer review, and no human expert or clinician reviews the analysis.

Corrections

If you believe something here is out of date or wrong, please tell us — see the About page for how to reach us. We'd rather fix an error than defend it.