[ ABOUT ]
About
Metatron is a grounded retrieval engine: answers come from documents it has read, carry the source they used, and refuse when nothing in the store is close enough.
What it is
Metatron does not answer from a language model’s memory. It looks a question up in what it has actually read, checks that the nearest passages are close enough to stand behind an answer, and only then hands them to a model to put into words. The model is the voice, not the authority.
Every answer carries a note on where it came from. When nothing is close enough, the engine says so — and the model is not asked.
The store
Everything the engine knows came from documents read into it. Each passage stays with the document and section it came from, which is what makes the note on an answer checkable.
At the time of writing, 18 October 2026, the store held 145,565 passages across 21 subject dictionaries — history, computer science, medical physiology, physics, and others. The full piece records the exact response and the count for every dictionary.
How one answer is made
The question is embedded, routed to a subject dictionary by the engine’s 13-node geometry, and matched against that dictionary’s passages on one measure: how near each one is to the question, as an angle on a sphere. The best match is checked against a grounding floor of 0.76.
If it clears, the model is handed the passages and writes; the engine attaches the source note. If it does not, the engine searches the web and measures what comes back; if that fails too, it refuses.
The refusal
The refusal is the feature, not a failure. On 18 October 2026 a question about a mission that does not exist came back with: “This isn’t in the store, and a web search found nothing to answer it from.” The engine had compared passages from four subject dictionaries (best 0.717 against the 0.76 floor) and run a web search (eight results, best 0.748 against an 0.800 floor). It declined without asking the model.
A calibration run on 13 October caught 51 of 60 out-of-corpus probes and produced no false declines on 27 answerable ones.
The writer
The language model is the voice, not the authority. It is handed passages and puts them into words; it does not decide what the store holds or where an answer came from. On the current box the writer is rented and is stopped between questions, so the first answer after a quiet spell takes longer. It can be replaced without changing what the store knows.
What it does not do
It does not learn at query time. The store changes only when documents are added, and a visitor cannot add one from the public internet. It can still misread a source: the store makes the evidence inspectable, not the interpretation infallible.
The house
Metatron Mind is a small, direct operation: a storefront, a publication and this engine. They share a palette, a type system and a set of rules, and they do not share a runtime, a database or a login.
- Ask Metatron — the AI. This page is about it.
- Metatron Mind — Shop — heavyweight garments, 200–430 GSM, printed to order and shipped direct. No account, no newsletter, no trackers.
- STACK — first-hand engineering essays with receipts: the engine, the box, the shop and the studio.
How we work
- Self-hosted and first-party. The shop and the publication load only their own files; the shop’s one third-party script is Stripe.js on checkout, where a card is actually charged. The AI calls out only where a job needs it: the writer and the image worker run on Modal, and X only if you connect an account.
- Receipts over adjectives. Every number we publish has a file, a command or a measurement behind it. Where we do not know, we say so.
- No ads, no analytics, no trackers. No advertising cookie, no fingerprinting, no session recording, no data sold.
- Separate stacks, one brand. A deploy to the shop cannot take the publication down, and an engine restart cannot change a price. The unity is in the surface: tokens, assets, language, links, policies.
Contact
Orders, shipping and returns: [email protected]. For a data request, use the same address with “privacy” in the subject and see the privacy page.
The full piece
This page is the short version. The full deep dive — the two live turns, the status and topology captures, the ladder’s code, and the commands to re-run all of it — is on STACK.