Explore the registry

Two pages read the whole registry rather than one record: "Explore" is every passport on Monad, and "Agents" is every generating agent that has signed one. Both are in the header, and both read the index live while the page is being built. Nothing on either is placed by hand.

Explore#

"The registry", newest first. The lede says where it comes from: "Every picture registered on Monad, newest first, read from the index while this page loads."

The page opens on the pictures. "Pictures" is the chip a bare /explore is on, so the first rows are ones a reader can look at, and the cards the repository's own scripts wrote sit behind the next chip, "Test images", rather than at the top of a judge's first screen. Nothing is hidden: "All" is the third chip and holds the whole registry.

Four counts sit beside the title: "Passports", "Generated", "Re-checked" and "Carry a statement on AI use". They count the whole index however the grid is filtered. The first and the last are the index's own running totals. The middle two are counted by asking for the matching ids, capped at a thousand, so a count that hit the cap prints as a number with a plus after it rather than pretending to be exact. The endpoint exposes no aggregate to ask instead. An index that does not answer the last counter leaves it out rather than printing a zero, which would read as a registry where nobody has spoken.

Under them, one line says how the registry divides: "17 pictures / 24 test images". That number is read the only way it can be, by fetching each pinned metadata document for its title, capped at the newest two hundred, and a read that hit the cap says so in the line.

The order, stated on the page#

One line in monospace over the grid. On "Pictures" and on "Test images" it reads "Newest first", because those two hold one kind of row each. On "All" and on every chip that can mix them it reads "Newest first, and seed test images sort last".

That second line is there because the registry holds cards the repository's own scripts wrote, titled Moolam seed 4, Moolam prove-it and Moolam Privy proof. They are not hidden and not dropped. They sort to the end, the page says so, and each one is labelled "Seed test image" on its plate.

The title is the only thing that separates them. Having a generator does not: the seed script signs every card it writes with the demo agent, so "has an agent" would call every seed card a test. A plate whose metadata will not load has no title to judge, and sorts with the test cards rather than above pictures a reader can actually see.

Inside a second, ties fall back to the id, so the same set always lands in the same order.

The filters#

Nine chips in two groups, a hairline between them. Which half of the registry is on the sheet: "Pictures", "Test images", "All". Then what the record says: "Generated", "Captured", "Edits", "Re-checked", "With manifest", "Disputed".

Each one is a real link carrying the whole state of the page, so a filtered view can be copied out of the address bar and sent to somebody else, and the browser's back button walks back through the filters you tried. /explore is the pictures, /explore?view=test the test images, /explore?view=all everything.

Beside them, a search box that matches a passport id. It is a form rather than a live filter, so typing does not fire a query per keystroke at the index. A link that carries a search term or an agent id and names no chip opens on "All": it is asking about records rather than about which of them are test cards, and the pictures chip would answer it with an empty grid.

Arriving from an agent's page adds one more chip, "Agent" and the id, with an x that clears it. A chip this build no longer has falls back to the same chip the rest of the address implies, instead of an error page, because a link someone shared should still open.

Narrowing by what the holder said about AI use#

Under the chips, a row labelled "AI use" with five of its own: "Any", "Open to AI use", "Not for AI training", "Ask first" and "Nothing stated". They carry the same words the passport page uses for the same statement, so the two screens can never name one picture two ways, and they combine with the chips above and with the search box.

A statement that answers each use on its own terms is summarised "mixed", and it appears under "Any" alone: no one of the four words above would be true of it. A picture nobody has spoken for is asked for by the flag rather than by the summary word, because the index writes none for silence and for a statement that named no use, and those are not the same answer.

The whole sheet is one indexed query. The index flattens the newest statement onto the passport itself, so nothing loops over the chain and nothing walks a statement history to draw this. The query is on the GraphQL page.

The two sets somebody outside the product actually asks for carry a line above the sheet saying what they are. On "Open to AI use": "Pictures whose holders say AI may train on them. Each statement is dated on Monad, and the passport page answers for any past day." On "Not for AI training": "Pictures whose holders have refused AI use, on the record, with the date."

An address that names an AI-use answer and no chip opens on "All", the same way a search term does: it is asking about statements rather than about which pictures are test cards.

The index cannot be asked for the two picture chips. It stores no title, and a title is the only thing that separates a real picture from a test card, so the page reads whole pages of rows from the index and sifts them with the same test the sort rule uses. That is why "Load more" under "Pictures" can read several pages of the index for one screen of plates, and why the cursor it carries points at the row the scan stopped on rather than at the last plate drawn.

A plate#

The pinned thumbnail and the title, with the kind in one word at the top left, "Generated", "Captured" or "Edited", and three marks at the top right.

A picture whose holder has written a statement carries the four marks of that statement under the kind, in the cold colours the passport page draws them in, and reads out as the statement and its date: "Not for AI training, stated 18 Sep 2026". A picture nobody has spoken for carries nothing there at all, because silence is drawn as silence. The line under the chips says what the four marks are.

A mark the registry does not support is a dim outline rather than absent, so a plate with nothing on it still reads as a plate with nothing on it. A picture the gateway will not hand over shows the seal and "No picture on the gateway"; a passport whose metadata carries no name reads "Untitled".

A sealed plate#

A passport whose picture was never published has no thumbnail to draw, so the plate draws its block hash instead: the same 16 by 16 grid its own page opens with. It carries a fourth mark, S, which reads out as "Sealed: this picture was never published", and the word "Sealed" beside the date, so the plate says what it is without anybody having to hover it. The line under the chips names the letter along with the other three.

It is never given "No picture on the gateway". That sentence belongs to a published picture the gateway would not serve, and the two states must not read the same.

The counts and the "17 pictures / 24 test images" line are unchanged. The index does not carry whether a passport is sealed, so the split cannot count sealed records separately without a change to the index's schema: a sealed record counts as a picture there, which is what it is.

Load more#

Twenty four plates a page. "Load more" asks for the next twenty four from where the last one ended, and the count under it says how many are on screen. The cursor is a pair, the registration time and the id, so two passports registered in the same second are never skipped between pages. A plate already on screen is dropped rather than drawn twice.

At the end: "That is every passport that matches." A filter nothing matches says "No passport matches this filter yet." rather than an empty grid.

Agents#

Every generating agent the index has seen sign a passport, busiest first. The order is the passport count, and nothing on the page is placed by hand.

The formula for the trust score is printed at the top, in full, next to a link to where it is worked out:

round(100 x (matched + 1) / (matched + mismatched + 2 x upheld + 2))

Each row carries the agent id, the name from its own ERC-8004 card, when the index first saw it sign, the trust score, and four counts: "Passports", "Re-checks matched", "Re-checks missed" and "Disputes upheld". "See its passports" opens Explore filtered to that agent.

The counts are the index's, because that is the only place they are added up. The name is the chain's, read from the card at the agent's tokenURI. An identity registry that will not answer costs the name and nothing else: the row still stands on its numbers.

With nothing recorded the page says "The index has not seen an agent sign a passport yet.", and an index that will not answer says that instead of drawing an empty table.

What one agent's own page shows is in Agents and generated images.

Sharing a passport#

Paste a passport link anywhere that unfurls links and the card is built from the record itself: the pinned picture, the title from its metadata, and one line under it, "Registered on Monad, block" and the block number. A record the index has not caught up with yet prints the chain's own registration date there instead, rather than saying it was registered and refusing to say where. A passport that pinned no name is called "Passport" and the short form of its id, and a picture the gateway will not serve falls back to the seal on a dark plate.

The title on the card and the title in the link preview come from the same function, so the picture and the text under it cannot drift apart.

The line in the footer#

Every page carries three live readings along the bottom: the verify service, the index and the chain. Each says "live" with what it is holding, the record count for the service and the block for the other two, or "unreachable", or how many blocks the index is behind.

They are read when the page is rendered, each capped at three seconds, and the page never waits on them. A red dot there means the thing is genuinely down rather than slow to paint.

Where to go next#

• Read a passport is one record and every block on it.
• Verify a copy answers which passport a file belongs to.
• Envio indexer is the index both pages read, and the trust score worked out in full.
• GraphQL is the same data, queryable yourself.