Identifiers truncated in this published copy. The record behind every row is retained in full.
123,751 launches recorded by this archive carried their launch document on one host, metadata.j7tracker.io. On 28 September we asked that host for 400 of those documents, chosen at random. It served 36 and answered 404 for 364. Every request got a definite answer, and the sample puts the share no longer served between roughly 88% and 94%, a conventional 95% interval, so on the order of 112,000 documents are gone from the only address their launches ever published. We then asked the public archive most likely to hold a second copy. Of 53 of the no-longer-served documents checked on 28 September, the Internet Archive's Wayback Machine held 0; asked again on 14 October for its complete index of that whole host rather than a sample, it holds 110 of these launch documents, 32 of which are documents this archive also holds, or 0.012% of the 266,784 held from the host by that date. A private copy somewhere is not ours to rule out, and a public one exists for those 32. The count is stated to the nearest thousand because it is a projection from a 400-document sample; every other figure on this page is a count.
The same host, asked for documents this archive already holds, on four dated occasions. "Served" means the bytes came back; a request that was rate limited or timed out is excluded from the denominator rather than counted as loss. The four samples were drawn under different definitions, stated per row, and the three smaller ones carry wide margins; the point survives all of that, because no ordering of 56, 0, 60, 9 fits a curve under any of the definitions.
| Date (2026) | Asked | Served | Share served | Where the sample came from |
|---|---|---|---|---|
| 10 September | 50 | 28 | 56% | first survival probe, random held documents on the host |
| 12 September | 75 | 0 | 0% | the published metadata-retention report, frozen that day |
| 20 September | 30 | 18 | 60% | spot check of held documents older than three days |
| 28 September | 400 | 36 | 9% | this exhibit's probe, random across all 123,751 held |
No retention curve fits those four readings. The share served is not a function of a document's age; it is a property of the host on the day you ask. An inspection on the 20th would say the documents exist. The same inspection on the 12th or the 28th would say they are gone. Both would be describing the weather. The stable fact is the capture: each document was fetched as its launch happened and hashed then, and the archive's copy does not change with the host's serving behavior.
The same day's archive-wide probe, 400 random documents across every host the archive holds.
| Population | Asked and answered | Still served |
|---|---|---|
| Content-addressed (ipfs) | 276 | 276 |
| Every other ordinary host, ten of them | 71 | 71 |
| Two hosts sampled at one document each | 2 | 0 |
| metadata.j7tracker.io | 49 | 6 |
Where the address is worked out from the file itself, the file survives: it can stop being hosted, but nobody can serve something different and still answer to the same address. An ordinary web address serves whatever its operator decides to serve, for as long as they decide to serve it. The third row is stated for completeness and proves nothing at one document per host. Of the 858,649 documents this archive holds, 577,587 are content-addressed and 123,751 live on the host in the bottom row, the largest single host in the archive after content-addressed storage, holding nearly three times the next.
The collector also keeps its failures: 27,551 launch documents it asked for at launch and never received. Of a random 400 of those re-asked on 28 September, 324 answered 404. Projected over the pool, about 22,300 recorded launches now have a published address that answers nothing: no host serves a document there, no replay of chain history contains one, and this archive never got one. For some, no document may ever have been served at all; a 404 today cannot tell an erased document from an address that never worked, and we do not guess which. About 2,100 still answer and are being fetched. Either way, the number is what capture at the moment of launch is for.
Revised the day of publication, before any citation: the control group's middle row first read 73 of 73 where the correct count is 71 of 71, with the two single-document hosts now their own row; the Wayback Machine check was added the same evening; and "gone at any price" in the last section became the narrower claim actually measured. The publication date is unchanged, and this note is the record of the change.
Revised again 14 October 2026, and this one corrects a statement rather than a count. The Wayback Machine figure published here on 28 September was "0 of 53 checked", and the page went on to say that a public second copy was something "we could not find". Both rested on a sample of 53 documents queried one at a time. Asked properly on 14 October - the Internet Archive's own index for the entire host, every path, which is a complete answer and not a sample - the Wayback Machine holds 110 of these launch documents, and 32 of them are documents this archive also holds. So a public second copy does exist, for 32 of the 266,784 documents held from this host by that date, or 0.012%. The earlier sentence was wrong and is withdrawn. The projection it sat beside is unaffected: an independent re-measurement the same day, 230 of the host's documents re-fetched at random, found 107 no longer served, a rate which implies about 117,000 gone and corroborates the ~112,000 above within 5%. Where both copies exist the Internet Archive's bytes hash identically to the copy captured here, and in all 23 overlapping cases examined this archive's capture was the earlier of the two. The publication date is unchanged and this note is the record of the change.
This page states what each request answered on the stated dates and does not call the host's conduct anything. A probe sees answers, not reasons: a retention policy, a storage migration, a lapsed account and a deliberate deletion are indistinguishable from outside, and we do not guess. Nothing here is a statement about any launch whose document the host carried, about the host's operator, or about any token; a document's absence from a host is never evidence about the token it described. We record what was served and when; what it means is the reader's question, taken up with the record in hand.