How this is written: The full account, including the uncomfortable part.

Pleroma began with my own practice and questions. I have lived with some of these teachings; others I am meeting here for the first time. The pages are drafted with AI from the primary sources, and nothing goes live until automated checks pass: in English posts, cited modern books and papers are looked up in public catalogues; on pages marked as checked, quotations are matched word for word at the source; internal links in the texts must lead to pages that exist; and the texts are screened for health claims. The line under each title says what its page had: “then checked: quotations matched word for word at the source” only where every quotation the checker found on the page was matched, and only the number of cited sources everywhere else.

None of this is a marketing claim. If a sentence on this page stops being true, the sentence changes.

How topics are chosen

Topics are not chosen by mood. A keyword engine collects what people actually type — the queries this site already surfaces for in Search Console, Bing's query data, autocomplete completions, People-Also-Ask phrasings, and questions harvested from public forums — scores them, and writes them into a queue.

A weekly plan takes rows off that queue. Each row carries the evidence that put it there: the query itself, and where it was seen. A subject nobody is asking about does not become a post, however interesting it is.

That is a search-first process and its cost is worth naming: it tilts the site toward questions that already have an audience. The counterweight is the source standard below.

How a draft is made

The prose is drafted by Claude, an AI model built by Anthropic, working from that brief and from the site's written standards — the voice, the structure, the source policy, the requirement that every post exist in both languages.

The Bulgarian is written as Bulgarian: its own sentences, its own register. It is not the English run through a translator.

The word “drafted” is carrying real weight in those sentences. There is no second pass by a person. That is the next section.

The source standard

A primary source is named with its edition and its date. A translation is named by its translator. A claim that rests on a modern reading of an ancient text says which modern reading, and when it was made.

The rule is that nothing is cited that was not opened. Where a claim is contested — a late attribution, a number with no controlled study behind it, a practice tradition dressed up as physiology — the post is required to say so rather than smooth it over.

That is the standard the drafts are written against. The corrections log is where the standard is checked in public.

Documented, Debated, Legendary

Desk pages trace one contested claim or one primary text back to its sources. Every claim on them carries one of three labels:

• Documented: a primary source we opened shows it.
• Debated: the sources, or the scholars reading them, disagree. Both sides are named.
• Legendary: the tradition says so, and no document we opened does.

A legendary claim is always attributed to the tradition that makes it (“according to its first chapter…”, “the tradition holds that…”). It is never stated in this site's own voice.

What runs before a post is published

Automated checks run before publication, plus the content build itself, without a person present. A check that fails the build stops publication; the ones that only report list what they find and stop nothing:

• Frontmatter — every required field present and valid against the content schema. Anything malformed fails the build.
• Glossary coverage — every lexicon term a post links to must have an entry in both languages, or the build fails.
• Bilingual pairing — an English post with no Bulgarian twin is reported by name. This one warns; it does not block.
• Citations, English posts — every cited work is looked up in public catalogues (Crossref, Open Library) by title, author and year, or its link must resolve. A citation that matches nothing fails the build. Ancient and classical texts with no catalogue record are skipped, not checked. Bulgarian posts, guides and lexicon entries do not go through this lookup.
• Quotations — on pages whose byline says “then checked: quotations matched word for word at the source”, each quotation the checker finds on the page (block quotes, text in quotation marks, the passages in source pop-ups and side-by-side texts) was fetched from the web page or PDF it names and matched word for word. A quotation that is not found there, or that was edited after its check, fails the build. Phrases under four words are not checked, and the checker finds most quotations, not provably every one. Where a quotation is our own translation, it is marked as one, and the check runs on the original wording at the source. Every other page with sources gives only their number.
• AI label — the templates for posts, lexicon entries and guides must print the AI-drafted label and its machine-readable marker, or the build fails.
• Claim labels — a page that labels its claims Documented, Debated or Legendary must have its quotations matched at the source, as its byline says, and the same labels in both languages, and a desk page must fill all its sections, or the build fails.
• Health wording — the texts are screened for health claims. A sentence the screen reads as promising an effect on the body or on health (set words, or a verb such as lowers or improves next to a body word such as blood pressure, sleep or anxiety) fails the build, unless it is attributed to its source, negated, or plainly meant in a spiritual sense. The screen knows set phrasings, not every possible claim.
• Internal links — every link inside a post, lexicon entry or guide must lead to a page that exists, or the build fails.
• Site figures — the numbers the site prints about itself (lexicon terms, Grimoire chapters, protocols) must match the files, or the build fails.
• Keyword contract — the post's primary phrase must appear in the title, the meta description, the slug, the opening lines and at least one heading, and the post must answer at least two real questions people ask. This one only reports; a post that misses it is still published.
• Style and overlap — machine-sounding phrasing, link text that names a different page, and posts that overlap an existing topic are listed for a later fix. These only report.
• The build — MDX that does not compile never reaches the site.

Read that list again for what is absent. The citation check can tell you a cited book exists, and the quotation check that a quoted sentence really is in its source; not one of them can tell you whether a claim is true.

No one reads a post before you do

Posts are not individually reviewed by a human before publication. There is no editor, no fact-checking pass, no approval step between the draft and the deploy. Five times a week the pipeline runs, the gates pass, and a post or a lexicon entry is live.

This was decided deliberately, with the alternative on the table. The alternative was fewer posts, each one read first. The choice was to keep the cadence and be honest about what the cadence costs.

What it costs is this: for most posts here, the first human reader is you. Errors are therefore not caught before publication. They are caught afterwards, by readers, and fixed in the open.

The stance is sourced fast, corrected publicly. Only the second half makes the first half defensible, which is why the corrections log is public, dated, and names who reported what.

What is never machine-written

Replies to readers. Emails to a person. Anything posted under this name in a community or a forum. If you write, a person writes back — not a model, not a template.

That is a rule rather than a preference. The one exception is mail that every subscriber receives alike: the welcome emails and the Sunday letter are AI-drafted and say so in a line at the top of every email (decided 2026-09-26, D46). Their text is screened for health claims, and the letter goes out only if all its links open; the quotations in its answer must be matched at the source, or the answer is left out. A reply to you is always written by a person.

Who is responsible

Stefan's Alkimia holds editorial responsibility for everything published here: the choice of topics, the standards the drafts are written against, the gates that run, the decision to publish without a review step, and every error that reaches the page.

The model drafts. The responsibility does not transfer.

Write to: pleroma@pleroma.guide

AI transparency — EU AI Act, Article 50

Article 50(4) of Regulation (EU) 2024/1689 requires that text published to inform the public on matters of public interest be disclosed as artificially generated or manipulated. It has applied since 2 August 2026.

The same paragraph carves out an exemption where the generated text has undergone human review or editorial control and a named person holds editorial responsibility for its publication. This site does not claim that exemption. There is no review step with which to claim it.

The statement above is a statement about the site, and it names a person rather than a machine.

Found an error?

Write to pleroma@pleroma.guide. The page gets fixed and the fix is recorded, dated, on the corrections log — with your name on it if you want it there.