The shortest path through what’s known. Not a generated answer, and not a verdict on who is right.
Zetlyn maps the open web as a graph of pages and the relations between them. For every page it crawls it records who it cites, whose wording it reuses verbatim, which pages name the same distinctive subject, which are copies of one another, and which apparently separate sites are really one publisher.
What that graph is for is the order you read in. A ranking hands you ten links and leaves you to discover that four of them are the same company and two are the same text. Zetlyn collapses those first, then plans a route: each step adds a publisher you have not heard from, or an aspect the earlier steps missed, or a citation path worth following — and it names which of those chose it. For an agent that pays per page fetched, a shorter route to the same answer is less context spent, a smaller bill and a result sooner.
Zetlyn is built in public and the only honest progress report is its size, so that is the roadmap. Each step is a corpus, not a date — and the figures below come from the same published snapshot that answered your last search, so you can check them against it.
Reading the current snapshot…
Whether two sources agree. That is a deliberate limit, not a missing feature. Zetlyn used to run a language model over pairs of claims and label them supports or contradicts. That judgement was the one thing in the whole system a reader could not check for themselves — and on real crawls it fired vanishingly rarely for what it cost. It was removed, and nothing was put in its place.
What you get instead is structure you can count and verify: six independent publishers cite this page, two of them reuse its exact wording, three of these “sources” are the same publisher. Deciding what that means is your job. It was always your job; Zetlyn no longer pretends otherwise.
Nor whether a page is worth reading. A route is not a recommendation. It says a page adds something the ones before it did not — a publisher, a subject, a citation — and which rule said so. Whether that addition is any good is the thing you are reading to find out.
/why?from=…&to=…&kind=…
about a relation between two pages and it answers with the rule, its version, the countable and how
to recompute the whole thing yourself — and where a relation rests on something you
cannot see by following the link, the interface offers show receipts, which asks
for you: the analytics fingerprint behind two different domains, the hash of the extracted text,
the count of shared 12-word runs, the rare subjects two pages have in common. For a plain citation
it offers nothing, because following the link is faster than reading a rule name. No confidence scores: a rule either fired or it
did not./route plans what to read, and takes a list of what you have
already read so a second call does not hand you the first call’s pages again./results is a finished answer: one entry per source, every reference
resolved to a url and a title, nothing to join and no ids to look up. Then
/route for the reading order, /record for one page in full,
/why for a single relation’s derivation, /traverse for the named
walks, and /graph when you want the raw edges between the results rather than
a finished list./context returns, per source, its summary, its subjects and the structure around it
— how many independent operators cite it, how many reuse its wording, how many of its
apparent sources turn out to be one publisher — with a stable record link and
no verdict. It is the one route that needs an API key
(Authorization: Bearer …), because it is metered per key; everything above is
open. A prompt gets countable structure and decides for itself.A polite crawler fetches public pages (see the crawler info). For each one it measures verifiable properties, computes an embedding, extracts named entities and takes two fingerprints of the text. It then probes the graph for pages sharing any of those — a link, a fingerprint, an entity, a publisher — and writes the page together with the relations it found. That is the whole pipeline: no second stage, no model deciding anything.
A few facts about a page depend on the corpus around it, and that corpus keeps changing. How distinctive an entity is, whether both ends of a relation are still indexed, whether two domains have since turned out to share an owner — all of these are worked out when you ask, never frozen into the stored relation. So the graph you are shown is current, not a snapshot of what happened to be true when the page was crawled.
The mark is a river through a circle: one path, chosen through everything else.
Zetlyn comes from zetetic — from the Greek ζητέω (zēteō), "to seek, to inquire." A zetetic method proceeds by investigation rather than assertion: the ancient skeptics were called Zetetics because they kept inquiring — testing a claim instead of accepting it on someone's authority. That is the posture of this project. Zetlyn hands you the sources and how they relate, to follow and check for yourself, rather than an answer to take on trust.
General: hello@zetlyn.com · Abuse / takedown: abuse@zetlyn.com · Security: security.txt