The stamp atlas that would rather say "I do not know"

I have been collecting stamps for over twenty years. That is somewhere between forty and sixty albums, thousands of small paper windows onto a place and a year, and until recently the only record of any of it was the albums themselves. Every tool I tried for cataloguing a collection this size wanted me to become a data entry clerk. A spreadsheet, a row per stamp, a lifetime of typing I was never going to finish.

The shortcut everyone reaches for now is worse than the tedium. Point a language model at a photo, let it write the description, and it will invent. It will tell you a stamp commemorates an event that happened after the stamp was printed, in a fluent and confident sentence, and you will believe it because it reads like every other caption you have seen.

Stamp Atlas is my answer to both problems. It is a public website that reads a collection as history and personality rather than as an inventory, and it shows only what it can honestly stand behind.

The one rule

The whole thing is built on a single rule: the atlas is never allowed to state more than it knows. Every number on the public site is one of three things. Confirmed, which means a person checked it. Candidate, which means a machine or a printed catalogue proposed it and nobody has confirmed it yet. Or unknown. When something is unknown, the site says so, in plain words, in the place where the number would otherwise go. "143 objects remain unidentified" is a line I am proud of, not one I am hiding.

That sounds obvious until you try to build it. The tempting design is a stamp record: one row per stamp, with columns for issuer, year, theme, and a description. The moment you have that row you have a second place for the truth to live. It ages the instant your understanding changes, and it is exactly the shape a language model is happy to fill with a plausible guess.

So there is no stamp record. A stamp's public page is composed fresh, every time, from a set of small dated attributed claims. If a claim is not there, the page says the thing is not yet known. It cannot show a fact it does not have, because there is nowhere for a phantom fact to hide. The honesty is not a policy I remember to follow. It is a property of where the data can and cannot come from.

It launched empty on purpose

On the day the site went live it said, honestly, that nothing had been photographed yet. It drew a compass rose and a blank page and told you to come back and watch it fill in. That is not a placeholder I was embarrassed by. Most of the map is still blank paper, and a real atlas of an unfinished survey should look like one.

It fills in as I photograph and identify, one album at a time. The public page holds a few kilobytes of facts and nothing else. No scans of the private albums, no prices, no storage location, none of the working notes. Just the aggregate the collection is willing to say out loud.

The sibling, and the twist

I built the coin version of this idea first. It is a small museum for ancient coins where a language model writes the story but can never invent a fact about my actual coins. I wrote up how that citation lock works in an earlier post. Stamp Atlas is the same spine pointed at a much larger and much older pile of paper.

Stamps pushed the idea further, because they come with a problem coins mostly do not. A modern stamp design is under copyright, so I can never show you the real thing. A later post is about how you illustrate a stamp you are legally not allowed to display. There is also a separate tool that reads one honest measurement off a stamp photo, its centering, which taught me that "cannot determine" is a first class answer long before I applied the same idea to a whole collection.

What it does not do

This is a slow project by construction. There is no "import my collection" button, because there is no honest way to import twenty years of paper in an afternoon. The map fills at the speed of a person looking carefully, which is to say it will never quite be finished.

I have made my peace with that. A collection that reads itself honestly and slowly is better than one that lies quickly. The next posts go under the hood: how a machine is allowed to help without being allowed to decide, how the whole thing survives being reprocessed forever from a single photograph of each page, and how the illustrations get drawn.