How this blog is verified | Evidence, checked by machine

This is a Japanese beauty blog. The articles are in Japanese, and they are written by AI.

This page is in English because the machinery is not in Japanese. The PubMed queries are already in English. You can run them yourself, right now, without reading a word of Japanese.

We are not asking you to trust us. We are handing you the tools to check.

What the machine checks, before anything is published

Every article is written by AI. We do not hide this.

Before an article goes live, a program queries the PubMed E-utilities API for every single PMID cited, and checks three things.

Does the paper exist?
A PMID that resolves to nothing is a fabricated citation. The build stops.
Does the title match?
A real PMID attached to a different paper's title is the most common form of citation fraud, and the hardest to spot by eye. The build stops.
Has the paper been retracted?
PubMed's pubtype field carries Retracted Publication. If a cited paper has been retracted and the article does not say so, the build stops.

If one paper fails, the site does not publish. Not the article — the site.

Articles published	11
Papers queried against PubMed	99
…of which industry-funded	29
Retracted papers found	3
PubMed queries handed to readers	48
…of which returned zero results	5
Last verified	2026-07-14

These numbers are written by the verification script itself (site/verify.mjs → site/verified.json). No human, and no AI agent, can type them by hand. There is no field in the article metadata where "verified: true" could be written. An article that has not passed the check cannot carry the receipt.

The part nobody else does: we hand you the search

Here is the problem with evidence-based beauty writing, including the good kind.

A writer says "I read thirty papers." The article cites two. What happened to the other twenty-eight? Which query was run? What came back? You cannot check. You can only trust.

Our core claim is almost always the same shape: "we looked for the source of this, and could not find it."

That claim was, for a long time, the one thing on this site with no evidence behind it. The existence of papers was machine-verified. The titles were machine-verified. The retractions were machine-verified. But "we looked" was just us saying so.

So now the machine records the search, as a side effect of running it. Not the agent — the search tool itself writes the log. Every article ends with the queries, the counts, and the dates:

Click it. It runs in your browser, against PubMed, right now.

And if a search returns nothing, that is the strongest thing this blog can publish — because you can confirm the nothing yourself.

Query	Results
`"skin fasting"[tiab]`	0
`(retinol[tiab] OR retinoid[tiab] OR retinoids[tiab]) AND ("vitamin C"[tiab] OR "ascorbic acid"[tiab]) AND (irritation[tiab] OR incompatibility[tiab] OR inactivation[tiab] OR concomitant[tiab] OR simultaneous[tiab]) AND ("randomized controlled trial"[pt] OR "clinical trial"[pt])`	0
`(PDRN[tiab] OR polydeoxyribonucleotide[tiab] OR polynucleotide[tiab] OR "salmon DNA"[tiab]) AND (topical[tiab] OR cream[tiab] OR serum[tiab] OR transdermal[tiab]) AND skin[tiab] AND ("randomized controlled trial"[pt] OR "clinical trial"[pt])`	0
`exosome[tiab] AND skin[tiab] AND (topical[tiab] OR serum[tiab] OR cream[tiab] OR cosmetic[tiab]) AND "randomized controlled trial"[pt]`	0
`exosome[tiab] AND (cosmetic[tiab] OR cosmeceutical[tiab]) AND "randomized controlled trial"[pt]`	0

Run them. The zero is the finding.

Two things this machinery caught, in its first week

1. Our own flagship article was imprecise

The article on the "28-day skin turnover cycle" said: "the PubMed search returns 0 results."

It returns six. All six are about rat lungs, rabbit vitreous humour, chicken bile acids, dog carnitine metabolism. None of them are about skin at all.

The conclusion held. The sentence did not. We rewrote it to show the query, the count, and what actually came back.

2. A retracted paper is still being cited in marketing

The first Phase III trial of polynucleotide (PDRN, "salmon DNA") filler was retracted in 2016 — for misstated conflict-of-interest and funding disclosures. A 2025 trial in the same field was retracted in 2026.

The machine found both. They are logged in an append-only corrections ledger, and the entry cannot be removed by deleting the citation — which is precisely the loophole the ledger exists to close.

What the machine does not check

This is the part that matters most, and it is the part that other "science-backed" beauty sites do not write.

Whether the papers are any good. A machine can confirm a paper exists. It cannot confirm the paper is right.
Whether we searched well. The log proves we ran a query. It does not prove we ran the right query. A badly written query returns zero results, and zero results are our strongest evidence. That is a real hazard, and we say so.
Medical accuracy. There is no doctor reviewing this. No dermatologist. No physician. We say so on every article.

We do not recommend products. We do not rank anything. We do not score anything. We have no certification mark.

That is not modesty. It is structural. A verdict is an asset, and assets attract buyers. Consumer Reports refused advertising for 86 years — and now charges manufacturers a fee to display the "CR Recommended" mark. Willpower does not hold that line. Not having the asset does.

Everything here is open

The articles, the verification script, the PubMed query logs, the corrections ledger, the figure data with its provenance — all of it is in a public repository.

There is nothing behind a paywall. There is nothing hidden.

The blog (Japanese) The source code

If you find a mistake in the machinery, or in an article, we want to know. That is not a courtesy — a blog whose entire product is "we checked" has no defence against being wrong except being corrected.