Corrections
These machines get things wrong. When that happens the page is corrected and the correction is written down, because a page that quietly changes its mind has taught you nothing about how much to trust it.
The account of what has actually been wrong lives in the method, in detail and without softening. This page is the other half: how to tell the studio about something, and what a machine has to be before it is worth suggesting.
A collection that is still moving
This collection has 111 machines built and 7 planned. It is not finished and it is not static: pages are corrected in place, and when a claim turns out to be wrong the page says so rather than quietly changing.
Readers and outside reviewers have raised 336 findings against these machines. 333 are closed and 3 were declined with a reason on the page. 335 carry the commit that fixed them, across 104 machines. The ones worth reading about are named in the method, which is where this studio keeps its account of being wrong.
What has been taken back
This site used to claim that an IBM 029 character is capped at three holes because a fourth would weaken the card enough to tear it in the reader or the feed rollers.
The three-hole maximum follows from the code: a character is a zone row plus a digit row. The 029's own MULT PCH key punches as many rows into a column as an operator likes, and the lace card is twelve holes in every column and a real object. Holes do remove material and a card cut too far through does tear, but that is not why the character set stops at three, and nothing cited establishes the causal version.
This site used to claim that a write acknowledged by a majority is permanent, and that nothing will make it un-happen, on the strength of quorum intersection alone.
Intersection proves reachability and nothing more: any two majorities of the same set share a member, so a later majority contains somebody who saw the write. Turning that into a write that cannot be undone needs rules the machine does not model -- terms, leaders, and which value a node may accept next. The page carried the caveat in one panel and the readout beneath it announced the bigger word anyway, which an outside reader caught on 2026-09-02.
Telling us something is wrong
Write to [email protected].
- The page you were on, and what it said. A quoted sentence is worth more than a description of one, because the wording is usually where the error is.
- What you expected instead, and why. A source is ideal and an argument is fine; a hunch is still worth sending, because a hunch that turns out right has been the start of several corrections here.
- Your browser, if the machine misbehaved rather than misspoke. A number that is wrong everywhere is a different defect from a control that only fails in one place.
Suggesting a machine
- It has to be a mechanism, not a topic. 'Compression' is a subject; 'the dictionary is never sent, and both ends build the same one out of the data' is a machine.
- There has to be something to compute. If the page would only describe, it belongs somewhere else: every machine here computes what it claims on the reader's own screen.
- It has to be sourceable. Every machine cites a primary document that a reader can reach, and the studio has turned down candidates whose sourcing was an account of the thing rather than the thing.
It also has to be one of three sizes, and the size is decided by how much of a system the thing is:
- Bit
- One idea, with nothing in it handing off to anything else. Lives on logicalart.us.
- Nibble
- One mechanism in three or four stages, each one reachable only through the one before. Lives on logicalart.us.
- Byte
- A complete machine you operate, on its own domain, with its own sources and honesty ledger.
What this page does not promise
Not every report becomes a change
Some findings are declined, and the count above says how many. A declined finding is not ignored: the reason is written on the machine it concerns, so you can disagree with the reasoning rather than wonder whether anyone read it. The commonest reason is that the page is deliberately modelling less than the real thing and already says so.
Suggesting a machine is not a queue you can join
The roster is small on purpose and most candidates are turned down, usually for sourcing rather than for interest. The studio has a bench of researched candidates it has not built. A suggestion is genuinely useful, and it is more useful with a document attached than with enthusiasm attached.
There is no comment thread, and that is deliberate
Corrections arrive privately and land on the page, where the next reader sees the result rather than the argument. That is slower and it loses something, and it is the trade this studio has made: the record of what was wrong is kept in the machine's own honesty ledger, which every page carries, rather than underneath it.