How this ledger checks ACBuy sheet links

The five readings

A reading describes one address at one moment. It is produced by requesting the address once and classifying the response, and the classification is the whole of the claim.

Why five, and why these five

Two readings would not be enough. A single yes-or-no would force a timeout, a policy block and a removed listing into the same bucket, and those three call for different actions: one means try later, one means the address is fine but unreadable from here, and one means the listing is gone. Separating them is what makes a snapshot useful.

The fifth reading exists because a record can be absent for reasons that have nothing to do with the address. A request that never completed is not evidence about the listing, and writing it down as a failure would be a claim the measurement does not support. Unrecorded is the honest label for that, and unrecorded results are excluded from headline figures rather than counted against them.

What goes into a sample

Each period draws a fresh sample rather than re-checking a fixed list. Addresses come from two places: wrapped addresses collected from publicly visible sheet pages, and source addresses drawn from a public directory listing. The two groups are counted separately because they answer different questions — one describes an unmaintained sheet, the other describes a maintained list.

The sample is not random in the statistical sense, and the page does not claim that it is. It is a fixed-size draw from publicly reachable material, chosen so that a reader can go and look at the same listings. That makes it reproducible without making it representative, and the difference matters when reading a percentage.

What happens to one address

1. The address is normalised and stored as a hostname plus a path. Fragments and anything that looks like a credential are rejected before anything is requested.
2. The hostname is resolved and loopback, private and link-local addresses are refused, so a submitted address cannot be used to reach something inside a network.
3. A single request is made with a short descriptive user agent and an eight-second limit. No cookies and no account credentials are sent.
4. The response status, and the hostname of any redirect target, are recorded. No page content is copied and no price or title from the response is stored.
5. The response is classified into one of the five readings, with the time of the request attached.

Requests run with at most two concurrent connections and a gap between calls. That ceiling is not caution for its own sake: pushing a source host harder produces a run of connection refusals which look like a finding and are not one.

What a reading cannot carry

A reading is not a stock notice. An address answering normally says nothing about whether the item is available, whether the size you want exists, or whether the price is the same as when the sheet was written.

It is not an authenticity check, a seller verification or a quality signal. This site does not assess whether goods are authorised, whether a batch is good, or whether a seller is reliable, and no part of the ledger should be read as implying otherwise.

It is not a forecast. An address recorded as reading normally can stop answering tomorrow, and one that returned a redirect can settle back to its original path. The date attached to a reading is part of the claim, not decoration: a reading without its date would be a very different and much weaker statement.

Limits that show up in practice

The ceiling on request rate has a visible cost. In the first period, 113 directory addresses reached the host after it had stopped accepting connections within the collection window. They are excluded from the table entirely rather than written down as unknown, which is why the published sample is smaller than the number of addresses the collection started with.

Shortened addresses are a second limit. They appear on the platform domain and are recognisable as shortened forms, but this site holds no verified expansion for them, so they are reported as read but not resolved. The same rule applies to source marks: only those observed in collected samples are mapped, and everything else is left unmapped rather than guessed.

A third limit is coverage. The sample is a selection of tens of addresses, not a register of everything a sheet might contain. Most product numbers will never appear in it, and a number that is absent from the sample has not failed anything.

Corrections and how they are handled

If a reading on this site does not match what you see, the correction route takes it. What makes a report actionable is the four things listed there: the page you were looking at, what you saw, when you looked, and where it can be checked independently.

Corrections that change a definition rather than a record move the method version. When that happens, earlier periods keep their original readings and are not recalculated: a record that quietly changes is worse than a record that is out of date.

What this page is not

This is not a compliance or safety statement. It describes a measurement and its limits. Nothing here should be read as a certification of any platform, listing or seller, and no platform has reviewed or approved this method.