This period
The window labelled 2026-W39 was published on 29 September 2026 and holds ninety-three records. Thirty-one of them are wrapped addresses collected from pages that were already public, and sixty-two are source addresses drawn from a public directory list. Both groups were read in the same stretch of days, under one set of rules, by one reader. Nothing about the split is decorative: it decides what can be claimed about wrappers and what can only be claimed about listings. A reader who knows which group a record came from already knows how far any conclusion about it will stretch.
The status column shows one value for the whole file. RESOLVES is recorded ninety-three times, with MOVED, GONE, SHIELDED and UNRECORDED all at zero for this period. A status here means the address answered when it was read, which is a narrow statement about a specific hour. It is not a stock notice, it is not a statement about a seller, and it says nothing about what will answer next week. It is a reading with a date on it, and the date is the only thing that makes it comparable with anything later.
What the period does not cover is written on the page as well. Every source marker observed in the sample pointed at one marketplace, so a single market reading sits behind every count. One hundred and thirteen further directory addresses were excluded rather than recorded as unknown, because the host stopped accepting connections during collection. No account, no cookie and no credential was used, and each request was given a fixed short wait before it was closed off. Two kinds of number are deliberately absent from that page: figures this site could not read during collection, and figures it could only have estimated.
Read the period from the sample IDs outwards rather than from the totals inwards. The weekly snapshots page is built to be read that way, one record at a time. Each record names an ID, a product number, a market, a status, and the day it was read. The checking method page carries the boundary language that sits behind those records: what counts as a reading, what counts as an exclusion, and why a refusal during collection is never written down as a disappearance. Those definitions travel with the numbers and belong beside them. A table published without its definitions is only a set of opinions with digits attached.
The period before
There is no earlier period to print. 2026-W39 is the first window this site has published, so the column that would hold a previous reading is empty, and it stays empty on purpose. Filling it with an estimate, a reconstruction, or a figure borrowed from somebody else would produce a row that looks exactly like a measurement while carrying none of the properties of one. An honest blank costs a reader one line of explanation. A fabricated comparison costs them a decision.
When a second window is published, that column will hold the same sample IDs read again, each with its own date. The point of holding the sample steady is that two readings of one address can then be placed side by side, and the difference between them is a fact rather than a suspicion. Comparison needs both dates. With one dated reading and one undated memory, a change cannot be separated from a different sampling frame, a different reader, or a different afternoon. Two readings taken a week apart under the same rules beat two readings taken by different hands in different months.
Backfilling was never on the table, because an earlier reading cannot be produced after the fact. A number typed into the table later would carry no date, no evidence class and no record of what answered, and it would sit among the measured rows as an equal. The value of the second period comes entirely from the discipline of the first one, which is why the first one refuses to invent a predecessor for itself. A gap in a published table is a statement, and it is a statement that can be checked later against the day the table was released.
What a reader should do with a first period is use it as a baseline and nothing more. Check the composition before the counts, notice that the whole file traces to one marketplace, and treat each figure as the output of a stated rule rather than a property of the file. The open questions page takes the same position where the public record runs out: the boundary is drawn and named rather than quietly rounded off. That page collects the places where an answer is missing instead of smoothing them over with a plausible range.
Every line that moved
No line moved this period, and the reason is arithmetic rather than editorial. A move requires two readings of one sample ID, an earlier status and a later one, and this is the first reading. The delta list is therefore empty for every one of the ninety-three records, including the thirty-one wrapped addresses whose wrappers failed while their rebuilt source addresses answered. Empty is the correct value here, not a missing one. A delta column filled with zeros that nobody measured would be worse than an empty column with an explanation under it.
A moved line will look different from a count. It will carry the same sample ID twice, the status recorded on each date, and a short note naming what was seen: a landing page that now points somewhere else, a listing that has been taken down, a verification step instead of a page, or a host that stopped answering during collection. The note matters as much as the status, because it is the part that can be argued with and checked again.
Per-ID deltas beat totals for a reason the current period already demonstrates. A total moves when the sampling frame moves, and one hundred and thirteen excluded addresses show how easily a count is produced by a decision about what to include. A delta tied to a sample ID cannot be produced by a change in the frame. It can only be produced by something happening at that address, on a date that both readings carry.
The lines most likely to move first are the ones already showing strain: source addresses on hosts that answered slowly, listings with little stock left, and pages that put a verification step in front of the product. Those are expectations and not forecasts, and the table will not be shaped to fit them. Whatever the second window records is what the second window records, including if the answer is that nothing moved at all.
Reasons the numbers moved
Four mechanisms can change a status without anything dramatic happening to a parcel. A seller removes a listing, and the address lands on GONE. A source address still rebuilds but arrives somewhere else, which is MOVED. A wall appears in front of the page, which is SHIELDED. A host refuses connections or runs past the wait, which is UNRECORDED. Each one is a different story and they should never be flattened into a single failure column.
Marketplace composition is the second reason a count can move without a link changing. Every record in this window traced back to one marketplace, so what looks like a site-wide rate is really a reading of one market on one set of days. A sheet built around other marketplaces would need its own window before a single number from it could be compared with anything here. Coverage is a property of the sample, and the sample is published with the counts.
The third reason is the rule about exclusions, which is easier to state than to remember. One hundred and thirteen addresses were left out of the table because the host stopped accepting connections, and the difference between an address that could not be read and an address that was not there is the difference between a missing reading and a finding. A table that hides the frame behind its totals cannot show that difference to anybody.
So the reading habit for the periods ahead is short. Compare sample IDs, not summaries. Read the composition before the totals. Keep the date attached to any number copied out of the table, because a number that loses its date becomes a claim about the present. By the fourth window this file will be able to say things the first one cannot, and it will say them one address at a time. That is slower than a summary and considerably harder to misread.