Report Accuracy and Public Data

Two people read this. One has bought a report or is deciding whether to. The other is the person a report is about. A report about a named individual is sold in one situation only: the buyer is the person it is about, and confirms that at checkout in a separate box that starts unticked. A report about somebody else is not offered, at any price, on any route, so for those two products the two readers above are the same person. Buying a report about yourself is not agreeing with what it says, and this notice is mostly about that. You may still be named in a report about a company without having been asked, because the subject of that report is the company; where such a report is in substance about you, clause 3.4 of If a Report Names You applies to it, and so does the export – the buyer can take that report out to a file and issue a link for it like any other, resting on the weighing in clause 5.7 there rather than on any agreement of yours. Both readers are addressed, and where a clause is for one of you only it says so. If a report names you and you want the route rather than the reasoning, start at If a Report Names You.

1. What a finding is, and what it is bound to

1.1 A report is made of the following, and this is the specification we undertake to meet. Every section of the rubric published at /intelligence-score for the report you bought, and for the combined package, which is two reports rather than one, every section of both rubrics. The rubric is versioned and the version your report was produced against is recorded on it, so what we undertook to deliver on the day you ordered is checkable afterwards rather than a number you have to take on trust. Six scored dimensions, each an integer from 0 to 100, scored against the rubric published at /intelligence-score. One headline score for each report, which is the plain unweighted arithmetic mean of that report's six dimension values, rounded once to the nearest integer, with no weighting, no curve and no adjustment of any kind. A minimum number of distinct sources, stated and counted in clause 9.2. Every finding bound to the evidence it came from and carrying a stated evidence strength.

1.2 What forms the description of what you buy. This notice is the complete statement of what a report will contain, how it is scored, how its sources are counted and what evidence stands behind a finding. Where another page we publish states a quantitative or performance claim about a report that this notice does not repeat – a number of sources, a length, a turnaround, or a degree of verification or cross-checking – that claim is withdrawn and does not form part of what you are buying. Where something we told you before you bought is more favourable to you than this notice and has not been withdrawn here, it still counts: this notice narrows nothing you were promised, and nothing in it reduces a right you have because of a public statement we made. If you bought after seeing a claim of that kind that we had not corrected, tell us and we will treat it as part of the description of your order and put it right under clause 7.6. So that this can be checked rather than argued about, we will keep an archived copy of this notice, of the product page and of the pricing page as each stood at the moment you ordered; where we cannot produce that record for your order, we work from the description you say you were given.

1.3 A finding is a single statement about the subject: a short title, a body, a tone of positive, neutral or risk, and at least one item of evidence. A finding is the smallest unit of a report that can be right or wrong, and it is the unit we correct.

1.4 A source is a retrieved artefact with a URL or a recorded query, a fetch timestamp and a content hash. Anything a model produces without retrieving it is not a source and cannot support a finding. That is the unit a finding is bound to; the floor in clause 9.2 is counted in a stricter unit, and 9.2 says which. Every finding will carry at least one evidence identifier, bound at the moment the statement is written rather than reconstructed afterwards, and a finding whose evidence does not resolve against the retrieved bundle will not reach you. The report will fail instead of shipping without it.

1.5 No person writes or checks a finding before it is delivered. The text of a report is produced by automated systems working to a published rubric, and there is no human review step before delivery. There is one afterwards, and it is named on the report itself so that a reader finds it without coming here: the person a report is about can contest any assessment in it – a finding, a dimension value, the headline score – and have a person re-examine it, on the route in clauses 7.2 to 7.4. That is the reason clauses 4, 7 and 8 exist, and it is the safeguard carrying the most weight now that a report can be exported to a file and read by somebody we never hear from. The AI Transparency and Limitations notice describes how a report is produced.

2. Two kinds of statement: attributed facts, and our own assessments

2.1 An attributed fact records what a named third party has published. What we are stating is that a named source said something, and we name the source in the sentence itself. Our record can be accurate as a record even where what the source said is false. That is not an answer to the person the statement is about, because repeating something false about somebody does them the same harm whoever published it first. We know that, and clauses 4.1 and 7.4 say what we do about it rather than leaving you to argue with us. Two limits follow. We do not publish an attributed fact about a natural person on a single source: clause 9.7 applies to it without exception. And where an attributed statement about a natural person would be seriously damaging to that person, it is not published as an attributed fact at all, because attribution does not make it safer to read.

2.2 An assessment is our own evaluative characterisation of the material printed beneath it. Every dimension value, the headline score, the band label, the verdict sentence, every section summary, and everything under risks, growth opportunities, strategic recommendations and the ninety-day plan is an assessment. Each is an opinion, formed by applying the published rubric to the evidence cited in the report, and it is written and presented as an opinion rather than as a statement of fact about the subject. The evidence it was formed from is printed alongside it, so that a reader can work from the same material and reach a different view. It is ours: if it is wrong we change it or take it out under clause 7.4, and we do not treat it as somebody else's error.

2.3 Every finding will carry which of the two it is. A report containing an untyped finding is not a report we deliver: the release checks described in the AI Transparency and Limitations notice test the finished draft for it, and a draft that fails is not delivered and is refunded in full and automatically under the Refund Policy. We do not write an assessment in the grammar of an attributed fact, and we do not write an attributed fact as though we had verified the underlying claim.

2.4 The distinction is not presentational. It decides what happens when you tell us something is wrong, because the two are corrected in different ways. Clause 7 sets that out.

3. What we read, and what we never read

3.1 We read what an ordinary browser can reach without credentials. We do not sign in to a publisher's site, we do not use anyone's account, we do not defeat a paywall, and we do not evade a rate limit. Our fetcher identifies itself by name, we honour the robots instructions a publisher sets for it, and where a publisher has excluded us the report will say the source was unavailable rather than quietly leaving a gap.

3.2 The one place we do use an account, and it is not a publisher's. For a report about a person we will also ask four AI assistants what they say about that person by name, and we will record the answers. Those assistants are reached through our own paid accounts with their providers, so this is the one step in collection that is not an anonymous public fetch, and it is the one point at which the subject's name deliberately leaves our systems as the subject of a question. We do it because what an assistant says about a named person is part of what a reader of this product wants measured, and we treat the answer as what one assistant said at one moment and nothing more, on the terms in clause 3.7. The providers are named, with the safeguard relied on for each, at /legal/subprocessors, and what happens to material such a question returns that we are not willing to hold is in the AI Transparency and Limitations notice and in clause 3.4 below.

3.3 We do not buy files about people, and we do not use data brokers or people-search aggregators as a source. We do not read private profiles and we read nothing behind a login. A maintained exclusion list of broker and aggregator domains will be applied to every fetch before a page is read, and we will add to that list whenever we find a source that should have been on it. Where a source is excluded, the report records that a source was withheld rather than leaving a silent gap. If you find material in a report that one of these rules should have kept out, write to us at the address in clause 7.2: we will remove it, tell you how it reached the report, and correct the list.

3.4 Where each exclusion bites, stated exactly. Two of the exclusions in clause 3.3 are decisions about where we send a request, so nothing of that kind is ever fetched: we send no request to a data broker or an aggregator, and we send no credentials to a login. The restriction in clause 3.5 works differently, and we say so rather than let one sentence cover both. Material of the kind named there is refused at the point evidence is written to our store, so that it is not retained, does not reach a model call and cannot become evidence. That control does not exist on the date of this notice. It is a condition of the first sale of any report whose subject is a natural person, and no such report will be sold before it is in service. An earlier design applied the rule only when a report was composed, which would have allowed such material to reach our evidence store and our model-call logs first; that design has been superseded, and we will not describe the control as better than it is.

3.5 Where the subject is a person, we read the public professional record and nothing else. Politics, personal life, health, litigation, criminal matters and temperament are outside what a report may contain, as are the special categories of personal data. This is a restriction on the product rather than a promise about our care, clause 3.4 says at which point it bites, and it is the reason a report about a person is a narrower document than a reader might expect.

3.6 Nothing a buyer tells us about a subject is treated as evidence. The context a buyer supplies with an order helps us find the right person; it never becomes a finding, is never quoted, and is never cited as a source.

3.7 What a class of source can support is limited, and a report is written to those limits.

4. When a public source is wrong

4.1 A statement can be an accurate record of the public record and still be false about you. A page can attribute an event to the wrong person of the same name, repeat somebody else's error, or be wrong on its own terms. We do not treat "we described the record accurately" as an answer to somebody who has told us the record is wrong about them.

4.2 Where a report names a person and that person tells us a finding about them is wrong, the finding will be suppressed first and examined afterwards, and it will stay suppressed until one of the three outcomes in clause 7.4 is reached. Suppression is not an admission that the finding is wrong. It is what we do while we find out, because leaving a disputed statement about a person in front of a reader while we look into it is the one outcome we are not prepared to defend. Suppression reaches the report and every view through every share link, from the next request. It does not reach a file exported before you wrote to us, and clause 7.7 says what we do about that instead of claiming a reach we do not have. The state that makes suppression possible is described in clause 8.1. It does not exist on the date of this notice, and no report will be sold before it does. Clause 13.1 of If a Report Names You states the same rule in the same terms.

4.3 One set of periods, whoever writes and whichever of our documents they read first. We acknowledge within one business day of your message reaching us. Where the subject is a person, suppression under clause 4.2 takes effect within three business days. We re-examine the evidence within fourteen days and give a substantive answer within twenty-one days. These are the periods in clause 13.5 of If a Report Names You, and where any document of ours states a different period for the same thing, the shorter one is the one you can hold us to. Where the law gives you a right to an answer within a stated period, that period applies as well and nothing here displaces it. The clock runs from when your message reaches us, not from when we finish establishing who you are. If we are going to miss one of these periods we tell you before it expires, why, and when to expect the answer.

4.4 Recording your position alongside a finding is an outcome we reach after examining the evidence, in the case described in clause 7.4. It is not our default and it is not what happens while a dispute is open.

4.5 We do not rewrite what a publisher said. Where the dispute is with the publisher rather than with our record of them, we say so and we tell you what we relied on. Where you are a natural person and the statement is seriously damaging, we remove our finding rather than restoring it with an annotation; removing our own statement is not rewriting what a publisher said, because the publisher's page is untouched and still says what it says. What we will not do is quietly alter our record of what we read.

5. Not professional advice

5.1 A report is not legal, financial, investment, tax, accounting, employment or any other form of professional advice, and reading one does not create a professional or advisory relationship with LeMans Labs.

5.2 The positive statement matters more than the denial. A report tells you what the public record shows and what we assess it to mean. It does not tell you what to do.

5.3 Some sections read like advice, so we should say what they are. Risks, growth opportunities, strategic recommendations and the ninety-day plan are options derived from the findings printed above them and from nothing else. They take no account of your finances, obligations, contracts or plans, because we have not been told any of it and do not ask. They are not tailored to your circumstances. The decision and its consequences are yours.

5.4 A report must not be the sole or determining basis for a decision with material consequences, and must never be the sole or determining basis for a decision about a person. It is one input, assembled quickly from public material, by a system that can be wrong.

5.5 A report is prepared for the buyer. It is not prepared for anyone else and nobody else should rely on it, and the export release makes that worth saying twice rather than once. A buyer whose report is about themselves may show it to whoever they judge fit, investors and a board included, and clause 8.3 of the Terms of Service says so. That is their decision about their own document; it is not us describing the report to the person they show it to. We told the buyer what this is – public material, read quickly, assessed by a system that can be wrong, on the limits in clause 5.4 – and we have told the reader over their shoulder nothing at all, except what the document itself carries on every page. For a report about a company, the people a buyer may show it to under that clause read it for the buyer's own decision and on the same restrictions, not for their own separate reliance. Supplying a report to a third party as diligence on which that third party is intended to rely is dealt with in the Acceptable Use Policy.

5.6 Nothing anywhere in this notice removes or limits the rights the law gives you as a consumer, including your rights where digital content does not conform to what was described to you. Where anything here could be read as narrowing them, it does not, and clause 7.6 states them.

6. Not a background check and not a consumer report

6.1 This restriction is taken as a term of sale before you pay, and repeated on what you receive. Before checkout you will be asked to confirm it as a separate, unticked statement under clause 9.2 of the Terms of Service, and the wording you were shown will be stored with your order. The same sentence will appear on every report we produce, whatever the product, above the score, on every view of a report through a share link, and on every page of any file exported from a report. On every page rather than once at the front, because a page is forwarded on its own and the reader who most needs this sentence is the one who never saw the checkout:

This report is not a consumer report. It must not be used to make decisions about employment, credit, insurance or housing. It is not produced to the accuracy standards those decisions require.

It appears on a company report as well as on a founder or executive report. A report about a sole trader, a named partnership or a company with a single named principal is a report about a person whichever product was bought, and we do not rely on the product name to tell us which case we are in.

6.2 A report is not a background check. In building a report we do not query criminal records, court records, credit files, licensing registers, sanctions or watch lists, or identity verification services, and nothing of that kind is a source. We do not confirm that a person is who they say they are. Nothing behind a login is read at all.

6.3 A report is not a screening product, not an identity check, not an audit and not a due diligence report. Decisions of the kind named in clause 6.1 are regulated, and the law requires a process for them that this product does not perform. That is a statement about what the product is, not a preference about how it is used.

6.4 The prohibited uses, the subjects a report may not be about, and what we do when we find out are in the Acceptable Use Policy. The quoted sentence in clause 6.1 is worded identically in that document and in the Terms of Service, so that no version of it can be played against another.

6.5 One check that runs the other way. Clause 6.2 says what a report is not built from. It is not a statement that we never consult a restricted-party list. Before a report is produced we will screen the subject you named, and you, against the sanctions and restricted-party lists that apply to us, and that check decides one thing only: whether we may lawfully produce the report at all. Nothing it returns is a finding, is quoted in a report, is cited as a source, appears anywhere in what you receive, or is disclosed to you. Where it prevents us from producing the report, the order is cancelled and refunded in full under the Refund Policy, and we may not be able to tell you why. Screening by our payment provider covers the person paying us and tells us nothing about the person a report is about, which is why this check is ours.

6.6 Where a report about a person is sold. Everywhere, with no territorial restriction. This section applies wherever a report is sold, and where the product is offered has never narrowed it. The condition of sale itself is in the Terms of Service.

7. If something is wrong: what happens

7.1 If you bought the report. Write to [email protected]. Tell us what is wrong and, if you have it, which finding. We re-examine the evidence bound to that finding. Every finding has evidence by construction, so there is always something to re-examine.

7.2 If a report names you. Write to [email protected]. If that address does not accept your message for any reason, write to [email protected] and say so: every period in clause 4.3 still runs from the day your first message arrived. You do not need an account, you do not need to have bought anything, there is no form, there is no fee and there is no particular form of words. The route, the identity checks and your wider rights are set out at If a Report Names You. Clause 4.2 applies from the moment your message arrives.

7.3 Re-examination has three outcomes, and we will tell you which one you got.

7.4 First: the evidence does not support the finding. We remove or amend it, regenerate the report, and tell the buyer that a correction was applied. Second: our record of what a named publisher said is accurate, and the dispute is with what that publisher published. We tell you which source it is so that you can take it up with the publisher, and we record your position against the finding and mark it disputed. Where the statement is seriously damaging to a person and that person has told us it is false, the finding is not restored to the report: our evidence can establish that the publisher said it and cannot establish that it is true, which is the case clause 8.4 governs. Third: we find no error. We explain why, cite what we relied on, and tell you how to object and how to complain. Where the subject is a person, the finding stays suppressed under clause 4.2 until one of the three is reached.

A fourth outcome exists for a different thing. The three above are about a finding; the headline score and every dimension value can be contested on the same route, by the person the report is about as well as by the buyer. A person re-examines the judgement standing behind the value. The arithmetic does not move, because it is published and anyone holding the document can recompute it – what moves is the judgement, and only on evidence. Where we do not move it, we say why, and the person's position is recorded against the value in the same way a disputed finding records one.

Disputing an assessment rather than a finding. The three outcomes above are about statements. A dimension value, the headline score, the band and the verdict can be disputed too, by the person the report is about, in the same words, to the same address and on the clocks in clause 4.3. A person re-examines it. Not the arithmetic, which does not move: the headline is the plain average of six printed values and anyone can check it by hand from the document. What is re-examined is the judgement underneath a value – whether the evidence cited supports the criterion it was scored against. Where it does not, the value is corrected and the report is regenerated. Where it does, the value stands and your position is recorded against it, so that the number and your disagreement with it are read together. Clause 13.4a of If a Report Names You is the same route written for you rather than for a buyer, and the route is printed on the report itself so that somebody holding a file exported months ago can find it without finding this page.

7.4a One vocabulary, five words. Suppressed means the statement is removed from the report as it is read, and from every view through every share link, from the next request after we act – we serve each of those views ourselves, so there is nothing to wait for and nobody to ask. It is not removed from a file exported before we acted, because that file is beyond us; clause 7.7 is where that is set out, and clause 8.8 of the Terms of Service is where the purchaser's obligation to delete it lives. Revoked is said of a share link and of nothing else: the link stops serving views from the next request. Withdrawn is said of our copy of a report: it is taken back from the buyer and their licence ends. Neither word is ever said of a file somebody has already exported, and if you find a document of ours using either that way, it is wrong and this clause governs. Disputed means the position put to us is recorded against the statement. Contested means both at once, and is the state described in clause 8.1. The Terms of Service and If a Report Names You use these three words in the same sense and in no other.

7.5 Every claim will be logged with a reference: what was claimed, which finding, what evidence was re-examined, the outcome, and the date. A correction that leaves no record is indistinguishable from one that was ignored. We will keep that register, and each version of a report as it was delivered, for the period stated for those classes in the Data Retention notice, so that a question about a correction can still be answered after the report itself has expired.

7.6 The buyer's remedies, in two tiers. The first tier is not conditional on anything. If a report does not match what we described to you before you bought it, you are entitled to have it put right, and if we cannot put it right within a reasonable time you are entitled to a reduction in price or to cancel and be refunded. That right comes from the law and not from us, and nothing in this notice, in our terms or in any other document we publish makes it depend on our agreeing with you. We do not decide whether our own product conformed. The second tier is narrow and discretionary, and is where our judgement legitimately sits: a finding that the evidence does support, which you dispute on other grounds. The Refund Policy is the only place our refund position is stated, and it sets out how a refund is made.

7.7 What a correction reaches, and what it cannot. A correction reaches the report itself and every view through every share link, from the next request after we make it. We can say that without qualification because we serve every one of those views and check the link each time; there is no address anyone holds that works on its own.

What it cannot reach is a file exported before the correction, or anything a purchaser or a person they showed it to wrote down, screenshotted or retyped, and we will not pretend otherwise. We do not say that we have recalled or destroyed an exported file. We cannot see it, and a company claiming to have destroyed it would be reporting the only thing nobody can check.

What we do instead: we regenerate the report, we revoke every share link we issued for it, we tell the purchaser in writing, and we ask them to confirm that they have done what the next sentence requires. Under clause 10.7 of the Terms of Service the purchaser must stop using and stop circulating the superseded version, must delete the copies they control, including any file they exported, and must tell anyone they gave it to that it has been superseded; and the recall covenant in those terms requires the purchaser, on our written request, to withdraw the report from anyone they supplied it to and not to supply it to anyone else. We can require that of our own customer. We cannot require it of somebody who is not our customer, and we do not claim to be able to reach a copy that has passed beyond the purchaser's hands. Where the correction was not caused by anything the purchaser did, taking those steps costs the purchaser nothing, and clause 7.8 governs the money.

Every report, and every page of every exported file, carries a report version and a generation timestamp, so a superseded version can be identified as superseded on its face, page by page, however far a page travelled. If you are the person the report is about, we tell you what we did, whether the purchaser confirmed, how many times the report had been opened and exported before your message reached us and on what dates, and how many share links were issued and when each was used; a recipient's name is information about another person, and clause 14.1 of If a Report Names You weighs it as it weighs the buyer's. And if you ask, we give you a signed statement of the correction that you can show to anyone who has seen the report.

7.8 What suppression decides, and what it does not. Removing or suppressing a contested finding is not by itself a defect in the report. This notice does not decide what you are owed: clause 7.6 states the entitlement that does not depend on our agreement, and the Refund Policy states the refund position and governs if any other page of ours differs from it. Two things follow, and both are ours to honour rather than yours to argue for. What we owe you about the size and contents of a report is measured as the report stood when it was delivered to you, so that nothing we owe you can ever be a reason to leave a disputed statement about a person in front of a reader. And where a suppression means the report no longer matches what we described to you before you bought it, clause 7.6 applies in full: where a whole section is lost, where a scored dimension is left unsupported, or where the headline score can no longer be computed from six printed values, we tell you so on our own initiative rather than wait for you to notice it. Where an objection requires us to withdraw a report altogether, the Refund Policy says what happens to the money.

8. Contested statements

8.1 Contested will be a state we can put a finding into without regenerating the report. It will remove the finding from the report as it is read and from every view through every share link, from the next request, and it will take effect in minutes rather than at the end of an examination. It reaches no file exported before it was set, in the same terms as clause 7.4a. It does not exist on the date of this notice, and clause 4.2 says what that means for when the first report may be sold.

8.2 A contested finding is neither an admission that we were wrong nor evidence that the person disputing it is. We do not treat it as either, and neither should a reader.

8.3 The buyer will see that a finding has been suppressed and that it is under examination. Where the report is about the buyer themselves, which is the only situation in which a report about a named individual is sold, that tells them nothing they did not already know. Where a person is named in a report about a company, it may tell the buyer that the person now knows, and we accept that consequence, because the alternative is to leave a statement in front of a reader after being told it is false. Anyone reading through a share link sees the suppressed report from the next request. Anyone holding a file exported earlier still sees the statement, which is the case clause 7.7 addresses, and the reason a suppression is paired with an obligation on the purchaser rather than presented as the whole of the answer.

8.4 Where a statement is contested and the evidence cannot settle it either way, it stays out. We do not restore a finding because a dispute is inconvenient to resolve, and we do not restore one because the buyer would prefer it back.

8.4a The one case in which a suppressed finding returns. Suppression under clause 4.2 begins on a claim that is plausible on its face and does not wait on our establishing who you are. In parallel we check that the person asking is the subject or is authorised to act for them, on the proportionate route in section 10 of If a Report Names You. A finding returns to the report only where that check shows the person asking was neither, and never because their objection failed on the evidence. If we restore a finding on that ground we tell you first, we tell you why, and we do not restore it inside seven days of telling you, so that you can put the position right or take it further. Declining to send an identity document is not a ground for restoration.

8.5 Contesting a finding does not suspend the ninety-day period for which a report remains available, and does not restart it. What it can do is add to it. Where we suppress a finding under clause 4.2 or clause 8.1 and the suppression runs for more than three business days, we extend the buyer's ninety-day window by the number of days it ran, because a buyer should not lose access time to a suppression they did not cause and could not prevent. We apply the extension ourselves and tell the buyer the new date; nobody has to ask for it. Clause 16.3 of the Terms of Service extends the same window on the same principle where something outside our control interrupts access.

9. Source counts, evidence strength, and what they mean

9.1 Beyond the floor in clause 9.2 we make no promise about how many sources a report will read, and we publish no typical or expected number. How much material exists about a subject is a fact about that subject's public record rather than a feature of the product, and a number presented as a feature would be a claim we cannot keep. The figures printed in the sample reports are part of the illustrative format described in section 10; they are not measurements, estimates or expected numbers, and clause 10.3 says what a sample shows and what it does not.

9.2 The floor, and the unit it is counted in. What we do commit to is a floor. A company report will read at least forty distinct sources and a founder report at least twenty-five; both floors apply to the combined package, each counted against its own part. For this floor a source is counted as a distinct registrable domain, after copies of the same text carried by several outlets have been collapsed together, with all of the subject's own pages counting as one however many of them we read. That is the strict unit and it is deliberate, because forty pages from three websites is not forty sources and we would rather fail a report than count it that way. The floor is a minimum and a term of the contract, not an estimate. Payment is taken before a report is produced, so a report that cannot reach its floor is not delivered and the whole amount you paid is refunded automatically, with no request and no decision by anyone, under the Refund Policy: you will see the charge and then the refund. Clause 12.3 of the AI Transparency and Limitations notice and clause 10.1 of the Terms of Service count this floor in the same unit, and clause 9.3 requires the report to print both figures so you can check it against the document rather than against us.

9.3 The report will print its own numbers, so it can be checked against itself rather than against a marketing page: the number of items read and, as a separate and always smaller figure, the number of distinct registrable domains behind them, which is the figure the floor in clause 9.2 is measured against; the date of the oldest and the freshest item; the spread of evidence strength across the findings; and the number of sources that were excluded because they could not be tied to the subject named in the order. The two counts measure different things, and clause 9.5 explains why the smaller one is the one that carries weight.

9.4 A high score on a thin record is a statement about a thin record. Where a report meets the floor in clause 9.2 and the public record behind it is still thin, the report will say so and name what was checked, rather than filling the gap.

9.5 Every finding will carry an evidence strength, computed from the evidence rather than asserted by a model, and printed under that name on screen and in the report data. Where any document of ours, any product screen or any earlier material uses the words "confidence", "confidence level" or "confidence label", it means this value and there is no other. High means three or more sources, at least two of them independent of the subject and of each other, with the freshest inside ninety days. Medium means two sources, or three or more that all trace back to the subject, with the freshest inside a year. Low means one source, or sources all older than a year, or sources that all trace to a single origin. Where the sources supporting a finding contradict each other, the evidence strength is reduced by one level and the contradiction is stated in the finding. Independence is counted by registrable domain after syndicated copies of the same text have been collapsed together, so four outlets running one wire story count once. These rules are the whole of the computation, and no other input affects the value.

9.6 What that is not. Evidence strength describes how much independent and recent material supports a statement. It is not a probability that the statement is true. Several sources can independently repeat the same wrong belief, and a well-evidenced finding can be false where the public record is uniformly false. The thresholds in clause 9.5 are a documented calibration rather than a measured error rate. We do not publish an accuracy rate for our findings, because we have not measured one, and a number we could not stand behind would be worth less than saying so. What we publish instead is checkable: every finding is bound to its evidence when it is written, the evidence is cited in the report, and clause 9.3 requires the report to print the counts the evidence strength was computed from.

9.7 A claim supported by exactly one source is never published as fact. It is either written as attribution, with the source named in the sentence and labelled as single source, or it is dropped. This applies to figures as much as to characterisations: a number that appears in one place is that place's number, not a measurement.

10. The samples on this site

10.1 The seven complete reports published at /samples carry this statement, which is reproduced here without softening:

These are illustrative demonstrations of the report format. They are not real assessments of the companies or people named, the measurements in them are synthetic, and they are not affiliated with or endorsed by any subject.

10.1a How that marking travels. The statement above will appear in the page header, beside every synthetic figure rather than once at the top, and in the structured data for the page. Each sample will record whether its text was produced by our pipeline or written by hand, because a demonstration that misdescribes its own provenance is the same error in the other direction. The permission granted to search engines and AI answer engines in the Intellectual Property notice extends to company samples only: a sample whose subject is a person is excluded from that permission, excluded in our robots file, served with directives instructing search engines not to index it, and kept out of our sitemap. A number attached to a person's name travels further and lasts longer than the sentence that qualifies it.

10.2 Every number in a sample was written to show the format: the dimension values, the headline scores, the source counts, the percentages and the captions describing a corpus. None is a measurement of anything, and none is a statement about any real person or organisation.

10.3 What a sample shows, and what it does not. A sample shows the format, the section structure and the rubric, and the way a finding is written and cited. It is not a representation about how many sources your report will read, how long it will be, or what it will say about any subject, and no figure printed in a sample is part of what you are buying: not the number of sources, the number of findings, the length of a section or of the document, the dimension values, the headline score or the band. Every figure of that kind is synthetic and is marked as synthetic beside the figure itself. What you are entitled to receive as to content, evidence and source volume is stated in clauses 1.1, 1.2 and 9.2 of this notice. Clause 2.5.1 of the Terms of Service and clause 4.2 of the Refund Policy say the same thing, so that no version of this can be played against another.

10.4 The sample data predates two fields a delivered report will carry. A sample can show a finding with no evidence line, and carries no evidence strength at all. A delivered report will carry both on every finding, without exception. The samples therefore demonstrate a weaker document than clause 1.4 requires, which is a limitation of the samples rather than of the product.

10.5 Any person or organisation named in a sample, or anyone acting for them, may have that sample removed by writing to [email protected]. We unpublish it the same day, without argument and without asking for a reason, and we confirm to you when it is done. The samples are replaceable and the dispute is not worth having.

11. The status of this notice

11.1 For a buyer. This notice forms part of your contract of sale under clause 2.3 of the Terms of Service. The version in force when you ordered governs your order, and section 17 of those terms governs how it changes.

11.2 For a person a report names. This notice is not a contract with you, and you are not asked to give anything for it. The commitments in sections 3 and 4, in clauses 7.2 to 7.5, in clause 7.7, in section 8 and in clause 10.5 are commitments we undertake to honour towards you, in addition to the rights the law already gives you. They do not replace those rights, they do not narrow them, and nothing here asks you to waive, limit or give up anything at all.

11.3 Changes. We may change this notice for the future, and a change takes effect when it is published with a new date. A change never applies to a message you have already sent us or to a report already delivered, and the version in force on the day your message reached us governs how we handle it. Previous versions stay published, with the dates each was in force, so that you can check which one applied to you.

11.4 Law and complaints. Section 20 of the Terms of Service states the governing law and the forum for a buyer, and what it does not take away from a consumer. If you are the person a report names, nothing in this notice or in that contract affects your right to complain to your own supervisory authority or to go to a court that is open to you; section 15 of If a Report Names You says where.

11.5 What our customer contract does not do. Nothing in our contract with a buyer limits the rights of a person a report names. Clause 12.6 of the Terms of Service says the same from the other side, and section 16 of If a Report Names You says it again.

---

This notice is part of the LeMans Labs legal set and should be read with the AI Transparency and Limitations notice, the Terms of Service, the Acceptable Use Policy, the Refund Policy, the Privacy Policy and, if a report names you, If a Report Names You. Who we are, how to reach us and where to send a formal notice are in section 1 and section 21 of the Terms of Service and in section 2 of the Privacy Policy.