The policy
Two things get logged here, with the date and the page.
- Something a reader could have acted on was wrong — logged with the previous wording and the corrected wording, so you can see exactly what changed.
- Something a reader should have been shown was not shown to them — logged with what was missing and who missed it. There is no previous wording to quote, because the words were right; they were just not in front of you.
Typos, formatting and rewording that does not change meaning are not logged — that would bury the entries that matter.
It was added on 21 August 2026, after an entry was written that this policy did not describe: a warning that was published, correct, and hidden by a layout rule on small screens, so the readers likeliest to need it were the ones who never saw it.
Writing that entry and leaving the policy describing only the first kind would have been the same failure the entry was about — a true description of a narrower thing than what actually exists. A warning nobody sees is not a warning, and a log that can only record wrong sentences cannot record that.
A site that revises itself silently is a site whose past claims cannot be checked. If you acted on something we said in March, you are entitled to find out in June that it was wrong, and to see exactly what changed.
It also disciplines us. An error that has to be written down in public is an error somebody thinks about twice before repeating.
The log
22 August 2026 — and the behaviour now matches the promise
Where: the question box, and privacy again. This is the second half of the entry below it, written the same day.
What changed: a question is now recorded only at the moment you are actually told this site cannot help. It used to be recorded as soon as our own guides declined, which was before the statutes had been tried.
So the original sentence is true again — questions that got an answer are not recorded — and it is true because of how it works rather than because of how it is described. A question that merely took too long is not recorded either; running out of time is not the same as having no answer.
One thing genuinely changed for you, not just internally. Your browser now does the sending, where before our server did. If you close the tab before the answer finishes, nothing is recorded at all. That is on the privacy page rather than left implied, because who sends a thing is part of what is being sent.
Logged separately from the correction below rather than folded into it. The wording was fixed first and the behaviour hours later, and collapsing those into one entry would describe a tidier morning than the one that happened. A reader checking what was true at a given hour should be able to tell.
22 August 2026 — the privacy page said answered questions are never recorded. They were.
Where: privacy, in the section on what the question box stores, and in the summary at the top of the page.
It said: “Questions that got an answer are never recorded.”
Why that was wrong: the recording happens the instant this site’s own guides decline a question. The statute lookup runs after that, as a second step, and it does not go back and undo the record. So a question the statutes answered fully — a deposit question, an eviction question — was written down as one nothing could answer. The sentence described the intended design rather than the built one, and nobody had read the resulting list to notice.
It now says that a question our own guides could not answer is recorded, even if the statutes then answered it, and that this is wider than the page previously claimed.
What is being done about it. The wording is the interim fix, not the resolution. The behaviour is being changed so that a question is recorded only when the reader is actually told this site cannot help — which is what the original sentence promised. When that ships, this page changes again and there will be a second entry here saying so.
Logged because it is a claim about the reader’s own words, on the page they would check to find out. It is the second time this section of this page has described a narrower version of what the question box does than the one that exists — the first was on 21 August, four paragraphs above this one. Both were found by looking at what the code does rather than by reading the page.
21 August 2026 — this page’s own policy did not describe one of its own entries
Where: the policy at the top of this page.
It said: “Any change that alters the meaning of something a reader could have acted on gets logged here, with the date, the page, the previous wording and the corrected wording.”
Why that was wrong: the entry immediately below this one has no previous wording and no corrected wording. Nothing on the site was wrong. A warning that was published and accurate was hidden by a layout rule on short screens, so a whole class of reader never saw it. That is worth logging and this policy did not admit it existed — it described a log that can only record wrong sentences.
It now says that two things get logged: something a reader could have acted on that was wrong, or something a reader should have been shown and was not.
Logged because it is this week’s error one level up. Every other entry here is a description that was true when written and quietly stopped covering the thing it described. So was this one. Writing an entry the policy above it does not describe, and leaving the policy alone, would have been the same failure the entry was about.
21 August 2026 — on a phone held sideways, the warning under the question box was not shown at all
Where: the line under the question box, on every page. Only on short screens — roughly, a phone in landscape. On every other screen it was there the whole time.
What was missing: that a language model writes the answers, that the question is sent in order to produce one, and the instruction not to type a cause number, a county or anybody’s name. That line is the only place any of that is said at the moment of typing.
Why it happened: a rule added to stop the panel overflowing a short screen hid the line to save the space. It did what it was written to do. Nobody noticed that the thing being saved was the warning, and the reader it was hidden from — somebody on a phone, probably in a hurry — is the one likeliest to need it.
It now moves rather than disappears: on those screens the same text is put into the opening message, where it is part of the conversation instead of furniture around it, and where a rule about panel height cannot remove it.
Logged although no sentence on this site was wrong, which is not what this log usually records. The wording was right and was published; some readers were simply never shown it, and a warning nobody sees is not a warning. It is here for the same reason as the “exactly one chapter” entry further down: a threshold that is only applied to errors that make the site look bad stops being a threshold.
21 August 2026 — “if you can name one” was a condition the question box does not impose
Where: start here and its Spanish counterpart, in the paragraph for readers whose question turns out to be about the law rather than the courtroom. The homepage carried a narrower version of the same description.
It said: “the question box on this page will look up a Texas statute for you if you can name one.”
Why that was wrong: the condition is not real. Naming a statute gets you that section’s text, which is what the sentence describes. But since 16 August the box has also taken a question that names no statute at all, found the chapters itself, and answered from the sections it fetched. A reader who did not know a citation was told, in effect, that there was nothing here for them — when in fact the box would have tried, and would have labelled what it found as statute rather than guide.
It now says that naming a statute gets you its text, and that if you cannot name one the box will find the sections itself and answer from those, labelled as statute because nobody here has read them. The homepage now says the same thing in its own words.
Logged because it points a reader away from help this site can actually give. The privacy entry below is the same failure in the other direction — that one overstated how carefully an answer was made, this one understated what the box would attempt. Both come from describing a tool that has since changed and not going back to the description.
21 August 2026 — the privacy page said answers came from our guides “and nothing else”
Where: privacy, in the section describing what the question box sends and what it writes from. Published, on the page a reader opens precisely because they are deciding how far to trust the box.
It said: “The answer you read is written by the model, but it is written from the text of the guides on this site and nothing else.”
Why that was wrong: true on the morning it was written, when the box could only answer from our guides. Later the same day the box was given a second source — sections of the Texas statutes, fetched live and read by nobody here first — and a paragraph describing exactly that was added two paragraphs above this sentence, in the same file, by the change that created the behaviour. The sentence was not updated. It then survived two further edits to the page and a “last reviewed” date being moved forward over it.
It now says that which text the model may write from depends on which of the two paths answered you, describes both, and says plainly that nobody here has read the statute sections before you see them.
Why nothing caught it. This site has a check that asks whether a page’s words are still the words that carried its review date. These words were — nothing edited them. It has a second check, a registry of claims that an addition elsewhere can falsify, and this sentence was not in it. Both checks were built in response to this exact failure, six days earlier, and neither fires when a true sentence is left alone while the world around it moves. The sentence stayed correct English about a fact that had changed four paragraphs away.
Logged because it is the most consequential version of this error the site has made so far. The entry below it about a false privacy claim was caught before anyone could read it; this one was published for five days. It describes how much to trust an answer, and it understated the answer’s reach in the safer-sounding direction, on the page that exists to be checked. A reader deciding whether to rely on something the box told them would have been reading a description of a narrower, more careful tool than the one that answered them.
16 August 2026 — a false privacy claim was found staged for publication, and deleted
Where: nowhere a reader could see it. That is the point of the entry.
It said: “This runs entirely in your browser: nothing you type is sent anywhere, stored, or logged, and nothing here learns from it.”
Why that was wrong: correct when written — the question box did once run entirely in the page. It stopped being correct when the box started sending questions to a language model. The sentence was never updated because it had quietly become unreachable: it sat in the site’s shared text file, no longer wired to anything, still describing the old behaviour.
Nothing was published and nobody read it. It has been deleted, along with ten other stranded strings, one of which still described the chatbot as only able to point you at a guide.
Logged because unreachable text is the worst place for a false claim to wait. Nobody proofreads it, no review catches it, and it is trusted on sight by whoever wires it up next — so the moment it becomes visible is the moment it is believed. What this site actually does with what you type is on privacy, and it is generated from one source now rather than restated by hand.
16 August 2026 — “exactly one chapter” stopped being true four paragraphs later
Where: the rules this site cites.
It said: “Of everything this site cites, exactly one chapter lives there: Government Code chapter 52, on court reporters.”
Why that was wrong: true when written. Later the same day the question box was given the ability to fetch any section of the 27 Texas civil codes, and a section further down this very page was added to describe it. The sentence and the thing that falsified it ended up four paragraphs apart, in one edit, and neither the person writing it nor a reviewer reading the page afterwards noticed.
It now says “of the rules and statutes these guides rely on”, and states that what the box returns is not part of the list.
Logged despite being harmless — it misdescribed this site’s own citations, not the law. It is here because “exactly one” is the most certain construction in the language, orientation is what that page is for, and shaving the threshold on a self-inflicted error is how the threshold stops meaning anything.
What makes it worth writing down is that no search of the text could have found it. The sentence was still perfectly good English about a fact that had quietly changed elsewhere. Both this site’s author and an outside reviewer produced this same error class within an hour of describing it — the failure survives people actively hunting for it, which is why there is now a registry of these claims rather than a resolution to be careful.
16 August 2026 — the question box described itself too narrowly
Where: the line under the question box, on every page.
It said: “Answers are written from this site’s own guides by a language model, so your question is sent to produce one.”
Why that was wrong: true when written, and then the box was given a second thing to do. It can now also answer civil-law questions we have no guide for by fetching Texas statute sections live and summarising them — text nobody here has read. The reply itself labels those clearly, and how we write this describes them at length, but the line every reader sees still promised something narrower and more checked than what they might be getting. How much to trust an answer is exactly the kind of thing somebody acts on.
It now says that answers come from our guides where we have one and otherwise from the statutes, fetched live, read by nobody first, and labelled that way in the reply.
Nothing about the answers changed, only the description of them. The labelling in the box was there before this correction and is unchanged.
16 August 2026 — the homepage stated the court-reporter rule flatly
Where: the homepage guide list.
It said: “no record unless somebody asks”.
Why that was wrong: too absolute, and the hedge already existed elsewhere. Guides and the guide itself both say a reporter takes the proceeding down on request and that if nobody asks there may be nothing to appeal from. A reader who saw only the homepage got a harder rule than the one this site actually stands behind.
It now reads “takes the record down on request — if nobody asks, there may be nothing to appeal from”, matching the other two.
They are the same shape as each other: a sentence that was accurate when it was written, a change somewhere else that made it inaccurate, and nothing connecting the two. That is the failure mode this log exists for, and we now have a build check that looks for one version of it.
Tell us about an error
Point at the page and the sentence. If you have the rule or statute to hand, better still, but you do not need it. “This does not match what happened in my court” is a useful report and we would rather have it than not.
Reports that something is wrong take priority over everything else, including writing new guides. A wrong page is worse than a missing one.
Two ways, whichever suits you. Email info@nosetoprose.org, or use the form below. The form is not more official than the email; it just asks the two questions we would otherwise have to write back for.
No cause number, no county, no judge’s name, and above all nothing about the people in your case, least of all a child. We do not need any of it to check a rule, and you are in litigation — the same instinct that makes Rule 21c redaction matter applies to a box on a website. Describe the page, not your case.
What we are most worried about getting wrong
So you know where to look hardest:
- Anything county-specific. Local rules and standing orders vary and we cannot track all 254 counties. If a page describes a practice your court does not follow, tell us and we will qualify it.
- Rule text that has changed since we checked. Every citation on this site carries the date it was verified, precisely so you can see how stale it is.
- Anything that reads as more certain than it is. If a page states something flatly that is actually contested or varies by judge, that is a real error even if the statement is defensible.
Last reviewed 22 August 2026.