Reports

A launch's own account of itself is not uniformly durable

Published 12 September 2026. The figures below were computed from the record on that date and are not updated afterwards: they are read from a file written when this report was published, not recomputed when this page is built. The live view is here.

A launch's on-chain record carries a uri pointing at an off-chain document: the name, symbol, description, socials and image the launch declared for itself. The chain keeps the pointer. It does not keep what the pointer returns.

Across 289,976 launches this archive holds a URI for, the document was re-fetched on 2026-09-12 and compared byte for byte against what was captured at launch. Documents do not decay at a uniform rate, and age is not what separates them - the host is. Content-addressed documents lost nothing. Three of the four largest self-hosting domains lost nothing. One did not.

A further 4,177 launches in this record declare an empty URI rather than a missing one. They are excluded from the table below, because a launch that declared no document has no host to serve it - not because nothing is known about them.

Every document sampled below is one this archive holds the bytes of, so each is known to have existed. That is what makes this a survival rate rather than an estimate. A document counts as served only if the fetch returned the same sha256.

Sampling 25 launches at random within each day, all of them documents this archive holds, metadata.j7tracker.io returned the document for every launch under two days old and HTTP 404 for every launch over three days old.

That host carries 42,902 launches in this record - 55.0% of every self-hosted metadata URI in it. This archive holds 18,980 of their documents.

It says nothing about why. The observation is an HTTP status code against a URL over time. This register records what a request returned and when. It does not infer a purpose or a policy, and none should be read into the tables above.

It is not established that anything was deleted. What is established is that the host returns 404. A document that is not served and a document that does not exist are different facts, and only the first was observed.

A 404 served only to us would look identical. Against that: the same host, from the same machine and in the same minutes, served every request for launches under two days old. That is evidence and not proof.

Every figure here has a ten-day ceiling. The archive begins on 2026-09-02, so the hosts that lost nothing lost nothing in ten days - which is not the same as durable, and this report should not be cited as saying it is.

Coverage began 2026-09-02 12:29 UTC and the figures were taken from the record built 2026-09-12 17:46 UTC. Launches before then were not watched. No rows are excluded for how they were discovered: a launch found late still declares a URI, and which host serves that URI is not affected by when this archive noticed the launch. The only exclusion is the 4,177 launches whose URI is the empty string, stated above.

Two halves, and they are checked differently. The population table is the query below, run against the public-domain file; it will return larger numbers than those above, because the archive has grown since publication.

The survival table cannot be re-derived from any file, because it is a measurement of what other people's servers were doing on the day. To repeat it: take any launch from the record that has both a meta_sha256 and a uri, fetch the URI, and compare the sha256 of what comes back against the stored one. Re-run it later and the answer moves AWAY from these figures rather than toward them, as documents that were still being served stop being served - the opposite of how the population table ages, and the whole subject of this report.

Since 13 September 2026 the record database is licensed and downloads with an API key. Text above written before that date is left as published.