Publishing Standards

Most publications have a style guide. This one has both. A gate, which is a program that runs on every deployment and refuses to publish a page that breaks a mechanical rule below, and a set of editorial rules that no pattern can state without refusing honest reporting along with the claim it is aimed at. Every rule below says which of the two holds it. Where a rule is mechanical, this is not a policy we promise to follow: it is a mechanism that makes the violation impossible to ship. Where it is editorial, it is held by the brief this publication writes to and by the automated review that tries to refute every draft before release. We publish the rules, and which mechanism holds each, so readers can hold us to them.

Language the gate refuses, by machine

Measured against the rules the build runs, on 2026-09-17. A page carrying any of these is refused whatever part of the system wrote it.

• Manufactured urgency, held as four literal phrasings: a claim that stock is running out, a count of how many units remain, an instruction to buy at once, and a claim that an item is moving quickly. The rule matches those four phrasings and not the idea, so urgency written another way passes the gate and is left to the review.
• A purity figure, meaning a number and a percent sign standing against the word purity or the word pure, with no method named within 90 characters of it. Naming the method beside the figure clears it, in the form "by HPLC" or "by mass spec": the rule looks for that wording, so naming a method without it does not clear the figure. Word order decides the rest. A percentage spelled out in words, one written after the word rather than before it, or one that does not stand against those two words, is not matched.
• Em dashes, anywhere in the copy.
• Twenty-two phrases that mark generated prose at a glance. They are listed in the rules file rather than here, because printing one on a published page would trip the rule that bans it.

Language the brief and the review hold, not the gate

Until 2026-09-16 a word list inside the gate refused human-outcome framing, benefit vocabulary and dosing language by machine. The owner removed it that day. The reason is worth stating on a page about mechanisms: the list could not tell a report of what a trial measured in its participants apart from a claim addressed to a reader, because both carry the same words, and it was refusing the first along with the second. These rules hold that line now, and they are editorial rather than mechanical:

• The subject of a sentence is the study, not the reader. A measurement is printed per compound, attributed to the publication that recorded it, in the past tense, with the study population as its subject. A figure with no study attached to it is a claim this publication would be making itself.
• Quantities, routes and frequencies appear only as attributes of a cited study. Never an instruction, never the second person, never a suggestion.
• Compounds are never ranked against each other by a measured outcome. Separate trials with different populations, endpoints and durations do not license that sentence, and it would be ours rather than any study's.

These three are held by the brief this publication writes to, by the automated adversarial review that reads every draft before release, and on the compound pages by the judgment of whoever compiles the record. A rule held that way is a different mechanism from a pattern in a file, and it is not a weaker rule: it is a rule no pattern can state without refusing honest reporting too. This publication covers a research-use-only supply chain, and nothing here describes or endorses human use of any compound.

Fabrication the gate refuses

• Invented bylines, credentials, or reviewer names, and stamps for an editorial process that does not exist here: a byline, a reviewer line, a doctoral or medical credential, a last-checked date. This publication has no named editorial staff yet, and the gate refuses any page that pretends otherwise.
• First-person claims of work not performed, in the shape of the word we followed by a verb such as audited or tested. A statistic with no source named behind it is refused by the review rather than by the gate, which carries no pattern for it.
• Links to pages that do not exist. Every internal link in the built site is checked on every deployment, and a dead one fails the whole build.

Disclosure the gate requires

Terminus Research and Banana Peppers LLC were formerly under common ownership. As of September 2026, Terminus Research LLC is independently owned and operated: it shares no ownership or financial interest with Banana Peppers, and no Banana Peppers personnel hold any editorial role here. Because of that history, Banana Peppers is held to the same criteria and the same process as every other supplier, nothing more favorable, and we disclose the prior affiliation so you can weigh it yourself.

Any page that names Banana Peppers cannot be built without that disclosure present on it. This page follows its own rule: the gate refused an earlier draft of this very page for explaining the requirement while leaving out the disclosure. The rule is structural. Forgetting it does not produce a page missing a disclosure. It produces no page at all. The full relationship is documented at How we are funded.

Discussion

Readers can respond to any analysis. Every response is reviewed before it appears, and a published one has passed the same language rules as an article. Most of that review is automated. A response that breaks the language rules is refused mechanically. A substantive comment that alleges nothing about a business is cleared automatically. A response claiming something checkable about a website is checked against that website before it appears. Every allegation about a named business or person goes to a person, always. So does anything the automated review is unsure about. We refuse responses that impersonate a business or a person, that make claims about human use, that name private individuals, or that assert something about a named company with nothing to check behind it. Those four are a person's decision at review rather than a pattern match. Handles are shown as submitted and are not verified identities, which the discussion says on its face.

A response that names something checkable is worth more than a comment: it goes to verification, and if it holds it becomes a dated entry in the ledger, credited on the response itself. Disagreement with our findings is welcome and gets published on the same terms as agreement.

Signing in

Responses can be sent with or without signing in. The site works the same either way. A reader who signs in with Google has their display name attached and their response marked verified. That tells other readers a real account stands behind it. A reader who does not is shown under whatever handle they typed, marked as unverified.

We do not keep accounts, passwords, or sessions, and signing in stores no email address. A sign-in is checked at the moment a response is submitted, then discarded. What remains is a one-way fingerprint, which lets a returning commenter be recognised as the same person without us holding their address. Signing in is deliberately optional. Readers in this field often have good reasons not to attach an identity to a comment about a supplier, and the reports worth reading most are often theirs.

What this means for a reader

Anything you can read on this site passed every mechanical rule above on the day it deployed, and was written and reviewed against the editorial ones. That is a narrower promise than it may sound. The gate checks wording, links, and the presence of disclosures. It cannot tell whether a fact is true, whether a certificate was read correctly, or whether a supplier changed its pages the day after we looked. Those are the ways we can still be wrong, and the corrections process is the commitment that covers them.

The articles are machine work at both ends. An AI system does the research and the writing, an automated adversarial review is the only review a draft now gets, and no person has to read an article before it goes out; an administrator can still veto one. That review leaves a record beside the article and the release step will not publish a draft without one; pieces published before September 5, 2026 predate the record and carry none. Every article carries a notice that says so, at the top and again at the foot, beside a Report control that reaches the administrators. The gate on this page runs over that output like any other, which makes the limit stated just above the one that matters most here: a machine-written claim about a named supplier can be worded correctly and still be wrong, and that Report control is how you tell us. The asymmetry is deliberate and worth saying out loud: a reader's allegation about a named business is held to a stricter check than our own machine-written one, and that is the reason the Report control exists.