How to invent an institution that holds together — and how to read the Foundry when it tells you yours does not.
Provenance ships with eight invented institutions. So does every other copy of Provenance that has ever been sold.
That is fine for as long as you are learning the system, and it stops being fine the moment you are selling. Two hundred people own the Brantwood Insect Cabinet. If two hundred people list the Brantwood death's-head hawkmoth, the buyer sees the same moth in eight shops and buys from whichever is cheapest — and the only way to be cheapest is to be poorer.
The eight sources are a teaching set. Yours is the shop.
A source of your own cannot be undercut, cannot be copied by someone who bought the same product as you, and cannot be searched for by a competitor who does not know it exists. It is the one asset in this system that is only yours.
Thirty-three fields, one block of eight sections, and twenty-five specimens. That is the whole of it. The Desk cannot tell your source from the eight it shipped with, because there is nothing to tell apart — they are the same object, filled in differently. The Foundry exists to fill it in without you spending a day on it.
An invented archive fails in one of two ways, and they look opposite.
It agrees with itself too neatly. Every date lines up, every number is round, every label says the same thing in the same words. Nothing real is that tidy, and a viewer feels it before they can say why.
Or it agrees with nothing. Fourteen handsome cards that have no relation to each other. This is the commoner failure and it is what a market prompt produces every time, because a prompt makes one image and has no memory of the last.
What holds a real archive together is a small number of facts that keep turning up in different places. Not repeated — recurring.
| A fact | Where it surfaces |
|---|---|
| Drawer XI holds the night-taken moths | on the drawer label, and again on a luggage tag two objects later |
| The camphor ran out | in a postcard, and again on a receipt |
| "At light" | in the ledger rows, and again in the description of the place |
| Too much moon on Tuesday | in a letter, and again as the gap between two ledger dates |
None of those is announced. Nobody explains them. They are simply there twice, and the second time is what makes the first one feel true.
The Foundry asks the model for four such facts and then checks that at least three of them actually recur. That check is the difference between a source and a nice paragraph.
Question five — how did it end — is the one that decides everything else, and it is the one people answer laziest.
An institution that closed when the buses came in 1931 has a last ledger page with an unfinished column. One that burned has scorched edges and a replacement register in a different hand. One that was absorbed into a bigger body has a second stamp over the first. One that simply dwindled has thirty years of the same handwriting getting worse.
Archives are made by endings. A place that is still open has no archive — it has a filing system.
They take five minutes. Each one is doing a specific job, and knowing the job makes the answer better.
| # | The question | What it decides |
|---|---|---|
| 1 | What was it | the vocabulary of every label. A depot says bogie; a wardrobe says skip. Say an institution, a trade, or a household — never a single person. One person is a biography, not an archive |
| 2 | Where | the weather, the dirt, and what condition things arrive in. Invent the place name. A real one drags real history in behind it |
| 3 | What it kept | the twenty-five specimens. If you cannot picture twenty-five different things, the source will be thin — this is the question to change |
| 4 | When it began | the typeface, the paper, the money, the abbreviations |
| 5 | How it ended | the stamp, the absence, the correction slips, the last page. See above |
| 6 | Whose hands | the handwriting, the initials in the ledger, the signature on the letter. One person who touched everything |
| 7 | Printed like | the stock, the ruling, the ink weight |
| 8 | Its emblem | the one mark that repeats and quietly makes fourteen objects a set |
A lighthouse, a tram depot and a seed-testing station ship with the Foundry so nobody is stopped by an empty form. Every buyer of the Foundry has the same three. They are a way in, not a source. Change questions 1, 2 and 5 at the very least, and it is no longer the same institution.
Leave a question empty and the model fills it. It will fill it plausibly and generically, which is the exact texture you bought this system to avoid. The Foundry tells you how many you have answered for this reason and no other.
The check is not a grade. It is a list of things that would have gone wrong later, found now instead.
A missing field does not stop generation. It prints. A card comes back with
the word undefined where the custodian's name should be, and by
then it is on a sheet you were about to sell.
| What it says | What happened | What to do |
|---|---|---|
| Not exactly twenty-five specimens | the commonest failure by far. The model counted badly, or merged two lines | add or cut lines by hand. It is thirty seconds and does not need a rerun |
| A specimen note is too short | a line says three-quarter view without saying of what | write what is being drawn. Vague notes make vague images |
| Blue in the palette | the model reached for a colour the core rules forbid | replace the hex with an earth colour. Blue fights the brown ink on every object |
| The corrected row does not exist | the correction points at a ledger line that was never written | change corrRow to name one of the five rows. This is the check that most often catches a genuinely incoherent source |
| Fewer than three facts recur | the model wrote eight sections that do not know about each other | rerun rather than patch. This one is not a typo, it is the archive failing to be an archive |
| Market words found | vintage, antique, aesthetic, ephemera | delete them. No archive ever described itself that way; only a seller does |
Patch anything mechanical — a count, a colour, a capital letter, a missing field. Rerun in a fresh chat when the failure is about coherence: facts that do not recur, or four or more failures at once. A second attempt usually fixes everything together, because the fault was the whole shape and not the details.
Three rules. The first two are law and manners; the third is craft.
No real business, no real institution, no real trademark, no living person, no named historical person. Invent the place, the firm and the people. If a name you choose turns out to be a real firm — and this happens, especially with short trade names — change it. The Foundry's prompt forbids this, and a model will still occasionally reach for something real.
These are invented documents. Sell them as reproductions, as printables, as collage papers, as journal kits. Never as antiques and never as genuine historical documents. The one thing that would destroy a shop built on this system is a buyer discovering that a "found archive" was sold to them as found.
The subtler rule, and the one that leaks. A depot that closed in 1931 has no biro, no zip codes, no telephone area code, no plastic, and no word that entered the language later. When something feels wrong about a finished piece and you cannot say what, this is usually it.
The Foundry invents institutions. It will also, if you point it there, build an archive around a name that matters to one particular person — and that is a different business with different economics.
| A shop listing | A commission | |
|---|---|---|
| What you sell | one archive, many times | one archive, once |
| What it takes | twenty minutes, then nothing | twenty minutes, plus a conversation |
| What it is worth | a few dollars a download | considerably more, because nobody else can make this one |
The work is the same work. The intake is the difference: instead of inventing a place, you ask the customer for one.
The fifth is the whole commission. It is the recurrence that makes the set land, and it is the reason the customer will not stop looking at it.
You are inventing a plausible archive in the manner of a family's history. You are not producing their history, and you are not producing evidence.
Say so in writing, in the listing and again in the delivery note, in plain words: these are invented documents made in a historical style; they are not genuine records and they are not proof of anything. A customer who understands that is delighted. A customer who does not will eventually show one to an archivist, and then to you.
The main handbook has a chapter like this and so does the Etsy Desk, for the same reason: a document that admits nothing cannot be trusted about anything.
| Claim | Standing |
|---|---|
| A source built here drives the Desk exactly as the eight built-in ones do | Verified 16/08/2026 in the browser. A Foundry source was installed, selected, given the right catalogue number, and produced complete prompts for four different objects with no missing values |
| The validator refuses a broken source | Verified. Three deliberately broken sources, each caught. The test also found two faults in the validator itself — including one where an unreadable palette was reported as containing no blue. Both fixed, and the second is why there is now a check that fails when the palette cannot be found at all |
| Five test institutions, written to this prompt, pass the check | Four of five, first attempt. The fifth — a village school's cabinet of miscellany, chosen as the hard case — failed on one specimen note too vague to draw from. The themes were written down before they were run |
| Twenty minutes from blank form to installed source | Not measured. It is the design target and it is what it took to build the test sources, but nobody unfamiliar with the system has been timed doing it |
| A source built in twenty minutes is as good as one built over days | No, and it is not claimed. The eight took days each. What the Foundry guarantees is structure — every field present, coherent, and checked. Depth is authorship, and a form cannot sell you that |
| That buyers want their own source | Judgement. It follows from every buyer receiving the same eight, and from what a crowded niche does to prices. It is reasoning, not a measured demand |
| What a source of your own will earn | Unknown and unknowable from here. No number in this document is a forecast |
The honest summary: the machinery is tested and the faults it found in itself are written down above. The promise that a source of your own is commercially worth having is an argument, and you should read it as one.