The method

How these documents were read

This site does two things most references on the subject do not: it publishes what a document fails to say, and it refuses to pick a winner when two documents disagree. Neither is defensible without the rule behind it, so both rules are here — with the counts they rest on, recomputed every time the site is built.

Four verdicts, and three of them are not answers

“The warranty says nothing about vinegar” is a different statement from “no manufacturer forbids vinegar”, and both are different from “nobody has read this one yet”. Every page here keeps those apart, and the words never change from column to column: they come from one controlled vocabulary, and a page that invented a synonym would be telling a reader two findings are different when they are the same.

Stated: The document answers this question. The manufacturer’s own sentence is printed underneath.
Not stated: The document was read in full and does not answer this question. That is a finding about this document, not about the manufacturer and not about the law: another document from the same brand may answer it, and the absence of a rule here is not permission.
Unreadable: The publisher’s own bytes are not in our hands — the link is dead, gated, parked, or serves an empty shell. Nothing is claimed about what the document says, because nobody here has seen it.
Unread: The document is in hand and this field has not been read yet. No page prints this state today; it is defined so that the day it appears it cannot be quietly rendered as one of the other three.

What was read, question by question

15 questions asked of 163 documents, over 172 entries — an entry being one document read for one product line. The table is the denominator behind every silence on this site. Where a column says a document is unread, no page anywhere here reports it as silent.

Every question this register asks, with how many entries state an answer, how many were read and say nothing, how many could not be read and how many are not yet read.
Question	Group	Stated	Not stated	Unreadable
Notice period	Claim	131	35	6
Who can claim	Claim	154	12	6
Evidence required	Claim	158	8	6
Removal before inspection	Claim	19	147	6
Transferable	Claim	146	20	6
Steam mops	Care	67	99	6
Vinegar	Care	23	143	6
Ammonia	Care	27	139	6
Robot mops	Care	0	166	6
Powered rotating heads	Care	98	68	6
Wet mopping	Care	92	74	6
Wax, polish and coatings	Care	102	64	6
Approved-cleaner condition	Care	141	25	6
Humidity range	Care	95	71	6
Cleaning machines as a class	Care	2	164	6

A silence is a claim about a search

When this site says a document does not answer a question, the sentence is only as good as the words looked for. So the search is not described in prose: it is a pattern, printed on the page that rests on it, and run at build time against tokens it claims to match and tokens it claims to miss. A pattern narrowed one day says so at the next build, and the gate fails.

Not a precaution against a hypothetical: two patterns had to be widened after being measured against the corpus, and the steam column carries its own and its test rows.

It reaches every column, as of 25 August 2026. All 15 questions on this site now carry a published pattern, and the 1235 silences behind them were each looked for in the publisher's own bytes. They did not always: the five claim columns had no pattern at all until 24 August, when transferability got one and five documents turned out not to be silent. Netting the last four the next day moved a cell the other way — a Shaw row had read a remedy tier as a notification period and printed a one-year deadline the document does not impose.

A silence is about one document, and nothing more. Not about the brand: another document from the same brand may answer. Not about the law. And absence is never permission — a document that does not forbid steam has not allowed it, and this site never renders one as the other.

What is deliberately not inferred

Every item here is a mistake that was available and was refused. They are listed because a reference is defined at least as much by what it will not say.

One product line’s rule, applied to another: The one error this reference cannot recover from. A brand’s warranty often opens with a chart whose rows are product lines and whose answers differ down the column, and one manufacturer keys its care rules to the warranty tier rather than to the floor: a line sold as waterproof is forbidden a mop that a line one rung below is allowed. So a rule is never lifted off its own line, and where a document answers more than one way inside itself, every answer is printed with the publisher’s own discriminator instead of being reduced to a verdict.
A manufacturer the document does not name: A retail brand is recorded with a maker only where its own document names one. Naming a single manufacturer behind a store brand is a guess dressed as a fact, and the guesses in circulation about the brands in this register are demonstrably wrong.
A document written for another country: A hard exclusion, not a ranking signal. One manufacturer’s European brochure permits steam on laminate and ranks well in search; another’s most findable full care text was an Australian distributor copy. Both would have produced a confident answer with no application to a floor bought in the United States.
A judgement about a brand: No brand here is called wrong, misleading or unsafe, and no cleaning product is called unsafe in this site’s own voice — a prohibition is always reported as the sentence of the document that prints it. Where two publishers disagree, both sentences are printed and neither is chosen. Nothing on any page is coloured for or against: colour on this site never carries a verdict.
A fitting claim, from a fitting manual: Installation manuals are read here for what they say about moisture, acclimatisation and the conditions a warranty depends on. The operator sells materials and does not fit floors, so nothing on this site offers, arranges or prices fitting work.

The quotations

2,908 verbatim sentences are printed across this site, each cut from its document rather than retyped, each next to the document it came from and the date that document was read. Short verbatim quotation with citation and a deep link is the whole arrangement: no manufacturer’s file is ever rehosted here, and every link goes to the publisher’s own copy.

Quotations are checked against the fetched bytes rather than against anybody’s memory. Two failure modes were measured on this corpus and both are guarded: a broken text layer, where one publisher’s PDFs drop the fl ligature so “flooring” arrives as “ooring”; and a sentence that exists in an extraction and not in the document, which is what a two-column page produces when it is read straight across — every word present, none of them adjacent. The second one was live here and is written up in full, including the reason the quote checker could not see it: the check and the defect were reading the same extraction.

One document in this register has no text layer at all. Its two quotations were read optically instead, which is a weaker guarantee than every other quotation here has, and the page that carries them says so.

Corrections

Before publication. The reading was audited in rounds, and 28 entries were changed as a result — a document state corrected from “fetched” to “blocked” after a direct re-check, a title pulled back to the one the page actually prints, a date removed from a field it did not belong in. 2 further entries were added during those rounds. The record marks them, so the count here cannot drift from what happened.

After publication. Nothing yet — this reference has not been published long enough to have corrected a published page, and saying so is more useful than an empty heading. When it happens, the correction is published with its date and its reason rather than applied silently, and the sentence that was wrong is quoted alongside the one that replaced it.

A correction is worth more than the mistake it fixes only if it is itself checked. A finding is a claim like any other: on this site a published sentence was once retracted on the strength of a finding that turned out to be wrong, and the refutation round brought the original sentence back. So a reported error is reproduced against the document before anything on a page changes.

Staying current

Manufacturers version these documents slowly — the stamps in this register run from 2015 to 2024 — so they are re-checked monthly rather than weekly, and the check is a fingerprint comparison rather than a re-reading: every document whose bytes we hold carries a SHA-256 on the source register, taken on August 23, 2026. A file replaced silently at the same address changes its fingerprint, which is exactly the change a reader cannot see and a hash can.

No schedule is promised. A cadence that cannot be kept produces a series with holes in it, and a series with holes is worse than none at all — it looks continuous. What is promised is narrower and keepable: the date every document was read is printed beside it, and it is never quietly refreshed.

One publisher cannot be watched this way at all. It serves its warranty text out of a database application rather than as a fixed file, so there is nothing at a stable address to hash; 6 addresses in this register serve no document to anybody who follows them, and they are listed with what each one actually returns.

Which of these rules a build enforces

A rule that lives only in a paragraph is an intention. These are the ones a machine checks before anything is published, and each of them has a deliberately broken fixture proving it can still fail — a check that cannot go red is not a check.

Every silence is published with the number of documents it was searched over.
Every quotation is found verbatim in the fetched bytes of the document it is credited to, and no quotation exceeds the length a short verbatim quotation can be.
Every published figure is recomputed by a second walk of the raw record that shares none of the page’s own method — a recount that calls the same function is not a second witness.
No entry whose own rules disagree is ever reduced to a single verdict, and no entry whose rules agree is marked as disagreeing.
Every published entry carries an address, a title and a reading date.
No prose characterises a brand, and the operator’s trading name appears on no surface — including in comments, structured data and the machine-readable copies.

What is not machine-checked, and is stated plainly for that reason: whether a reading is correct. A gate can prove a sentence is in a document and cannot prove it answers the question it was filed under. That judgement is a person’s, and every entry names the document so it can be checked by anybody.

What this page does not know

Whether a prohibition and an exclusion are the same act. “Do not use a steam mop” and “damage from a steam mop is not covered” both end a claim, so this register files them together and says so. They are not the same thing in law.
How much of the market these 33 brands are. No free authoritative source for installed base was found, so coverage on this site is published as a count of brands and documents and never as a percentage of anything.
What is behind an account or an email gate. Only ungated addresses are collected; two manufacturers are absent for that reason alone.
Which document governs your floor. The one that came with it does. This site can show you every reading a brand’s documents carry and cannot tell you which of them was current the day your floor was sold.

The documents themselves, with their fingerprints: the source register. Who publishes this and what they sell: about this reference.