What we refused to publish
This site promises that no number reaches the shelf without a measurement behind it. That promise cannot be proved by showing you what we published — it can only be proved by showing what we took out. Every entry below was really written somewhere here, and was really removed.
These are not bugs. A bug fixed is hygiene; a claim deleted because we could not stand behind it is a decision, and only the second kind is on this page.
What went, one by one
A fifteen-million-dollar market figure
was on: the home page
Why it went: It was not ours, we could not check it, and it sat next to the word 'measured' on a site whose whole claim is that we measure things. A reader could reasonably have thought we had run that number ourselves.
What is there now: There is no market-size figure on this site. We do not have one worth publishing, and saying nothing is the honest version of that.
A page count written from memory
was on: the guide's own page
Why it went: It said seven pages. The file had eight. Nobody lied - someone wrote down a number instead of opening the file, which is how almost every wrong number gets published.
What is there now: The page count is counted out of the PDF every time the page is built. It cannot drift again, because nobody types it any more.
A denominator that turned a true number into a false impression
was on: the shelf counters
Why it went: 'We benchmarked 7, of all 437' is arithmetically true and reads as a confession. Most of those 437 are other people's free tools we collected and never claimed to have tested - putting a denominator there claimed coverage on our behalf that we had never claimed ourselves.
What is there now: No denominator. The counters say what we ran, what we put a clock on, and what is only collected - three separate facts instead of one misleading ratio.
Speeds for tools we ran but never timed
was on: dozens of cards on the shelf
Why it went: We have run far more than we have measured. It would have been easy to put a number next to each one from memory, or to borrow a figure from a neighbouring tool because the model name matched.
What is there now: Those cards carry a verdict and no number, and each one says why the number is missing. A verdict stays true on any machine; a speed taken on ours does not.
A payment method list that would have looked complete
was on: the machine-facing API
Why it went: An agent asking how to buy gets an HTTP 402 with an empty `accepts` list. We could have filled it with something plausible. An agent that tries a payment path that does not work loses more than one that is told the truth up front.
What is there now: The list is still empty, and the reply says so in words: there is no method an agent can complete on its own here yet.
Prices presented as settled
was on: the products we list
Why it went: The prices on this site were proposed by an assistant, not decided by a human yet. Presenting a proposal as a decision is a small lie that becomes a large one the moment someone plans around it.
What is there now: They are marked as proposals in our own data, and the deploy refuses to go quietly: it prints a warning every single time until a human confirms them.
A blockchain we did not need
was on: a proposal for signed certificates and tradeable tokens
Why it went: Of five ideas involving a chain, four did not survive contact with the question 'what does this prove that a signature does not?'. A chain proves who wrote what and when. It does not prove the thing is true, and true is the part this market is about.
What is there now: Ownership, provenance and transfers are kept in the open without a wallet, a gas fee or an account. If agents actually trade, the case for real settlement becomes a measurement instead of a bet.
Questions people actually ask
Why publish your own deleted claims?
Because the promise on this site - that no number reaches the shelf without a measurement - cannot be proved by showing what we published. It can only be proved by showing what we took out, and why. Anyone can promise rigour.
Are these bugs?
No, and the difference matters. A bug fixed is hygiene. A claim deleted because we could not stand behind it is a decision, and this page is only about the second kind.
Why do so many tools have no speed on them?
Because we ran 74 of them without a stopwatch. We could have written a number from memory for each. Each of those cards says what we found and says plainly why there is no figure - a verdict stays true on any machine, a speed taken on ours does not.
Can an agent pay you automatically?
Not yet, and the API says so rather than offering a path that does not work. The `accepts` list in our 402 response is empty on purpose.
The same habit runs through the shelf, where every tool says whether we measured it, ran it, or only weighed it on paper — and through the licence traps, which is the same idea pointed at other people's models instead of our own pages.