AI Transparency and Limitations

Version: 1.0. Superseded versions, and the date each was replaced, are listed in clause 12.5.

---

0. Status of this notice

0.1 From the effective date above, this notice is in force. It forms part of the contract of sale under the Terms of Service, and the version in force when an order was placed governs that order for the whole life of the report.

0.2 It is a description and a limit at the same time. Sections 1 to 5, 8, 11 and 12 describe what the product is, and form part of the description a buyer is entitled to receive. Sections 6, 7, 9 and 10 state what the product is not, what it is known to get wrong, and what we do not promise. Both halves take effect together, from the same date, on the same footing. Neither is commentary and neither should be read without the other.

0.3 Where a sentence says "will". This site is a marketing site. The checkout, the account, the pipeline, the report and the delivery email do not exist yet, as the Sub-processor list also records. A sentence written with "will" therefore describes something that does not run yet rather than something that does. Where such a sentence describes a control we treat as a condition of sale, we do not sell the product it applies to until it is true, and every control of that kind is listed, with the clause it belongs to, in clause 12.7.

1. The disclosure, in full

1.1 LeMans Labs OÜ, a private limited company registered in the Republic of Estonia (Commercial Register code 16872044), registered address Valukoja 8/1, 11415 Tallinn, Estonia, is the provider of the automated system described here. There is no separate legal notice page: full provider identification, our contact details and the address for notices are in the Terms of Service and the Privacy Policy.

1.2 One wording carries the disclosure everywhere, and it reaches the reader before the reader is asked for money. It will appear on every product page above the price, at checkout immediately above the pay control rather than behind a link, in the report header above the score, on any shared view of a report, on the samples index, and here:

This report is produced by automated systems that read publicly available information and score it against a published rubric. It is not reviewed by a human before delivery. Every finding carries the evidence it came from and a measure of how much independent evidence supports it, which we call evidence strength and which is not a probability that the finding is true. Automated analysis can be wrong, and where the public record is thin the report says so rather than filling the gap.

1.3 Why the two pre-purchase placements are not optional. That a report is produced by automated systems, that no person at LeMans Labs reads it before it reaches the buyer, and that it can be wrong are the main characteristics of what is being sold. A buyer who learns them only after paying has not been told them. Checkout will not open until both pre-purchase placements are live and asserted by the test in clause 1.6, the confirmation we send on a durable medium will repeat this wording, and we will record what a buyer was shown, as shown, with the time, so the placement can be checked rather than asserted.

1.4 Short form, for the delivery email and the foot of a shared view:

Automated analysis of public information. Every finding carries its evidence and how much of it there is.

1.5 Three elements of that wording are fixed and will not be softened in any later version: that no human reviews the output before delivery; that findings carry the evidence they came from; and that the system can be wrong, stated without hedging. If our position on any of the three ever changes, that is a new notice under a new name, not a revision of this one. Clause 12.5 says how this is secured.

1.6 The disclosure will be held as a single exported string and rendered from that one place, so the wording cannot drift between a product page, a checkout screen, a report and this page. A surface listed in clause 1.2 that renders without it will fail its build rather than render silently, and an automated test will assert that every such surface carries it. None of that runs today, because none of those surfaces exists. The string, the render gate and the test will exist before the first report is sold, and clause 12.5 will record the release in which they did.

2. How a report is produced

2.1 Before anything is collected: who a report may be about. A report about a natural person is produced in one of two situations only, and the check is automated in both.

A buyer's statement that they have the subject's permission is not enough on its own. It is kept, it is a warranty, and it carries the consequences set out in the Terms of Service, but consent has to be given to us and has to be something we can demonstrate, and a third party's assertion that permission exists is not a demonstration that it does. No person at LeMans Labs is involved at any point in this step: the request, the wait, the cancellation and the refund are all automated.

One consequence runs through the rest of this notice and is worth stating early. A person a report is about has heard from us before the report exists. The routes in clause 7.9, in section 9 and in If a Report Names You therefore reach someone who already knows we exist and already knows how to reach us.

(/legal/privacy) and [If a Report Names You](/legal/report-subjects) word for word.]

2.2 A report is produced in four stages, and those four are what a buyer will see named while it runs: collecting, analysing, scoring, composing. Nothing is delivered until all four have finished and the checks in clause 4.3 have passed.

2.3 Collection reads only what an ordinary browser can reach without credentials. We do not log in, we do not pass paywalls, we use no private profiles and we buy no data from data brokers. Our fetcher will identify itself by name and obey robots.txt for that name; a page we are asked not to read will be recorded as unread rather than quietly omitted. These are binding rules on the product, not descriptions of a capability.

2.4 For a report about a person, collection is confined to the public professional record. Politics, personal life, health, litigation, temperament, and any material revealing a special category of personal data, are outside what may be collected at all rather than merely outside what may be printed. Making that true requires a control at the point of retrieval rather than at the point of composition, and clause 3.7 states plainly where that control stands today and what it gates.

2.5 Some of the work will be done by code and some by a language-model judgement. The division matters, because only one of the two can be wrong in an interesting way.

2.6 What a finding is, how an attributed fact differs from our own assessment, and what happens when a public source turns out to be wrong are in Report Accuracy and Public Data. This notice covers the machinery.

3. What the models do, and what they do not do

3.1 The models are general-purpose language models built and operated by third parties. We build no models, we train none, and we fine-tune nothing on anything a buyer gives us.

3.2 A model will not be able to produce the headline score. The headline is computed in code as the plain mean of six dimension values and no model is asked for it, which is what makes the arithmetic in section 5 checkable rather than asserted.

3.3 A model will not be able to invent a source. Every finding must cite an evidence item the collection stage actually retrieved, identified by a reference a model has no way to guess; a finding citing anything else is discarded before the report is assembled. For the same reason a model cannot introduce a link or a citation of its own.

3.4 A model will not be able to add a claim while writing. Composition receives findings that already exist and may only phrase them. A figure in a finding must match a figure extracted from the evidence it cites, or it is removed and the sentence rewritten without it.

3.5 The models that read evidence will have no tools, no network access and no ability to act. They read text and return structured output, and that is the whole of their authority.

3.6 A model's own memory is never a source of fact about the subject of a report. Where we ask an assistant model about a subject we are measuring what that assistant says, not treating the answer as true: an assistant answer is evidence of the answer and nothing more. Clause 11.3 explains what that measurement is and is not.

3.7 Material we are not willing to hold at all. Asking an assistant who a named person is can return material that reveals or implies political opinions, religious or philosophical belief, trade union membership, racial or ethnic origin, health, sex life or sexual orientation. Our position is that no condition available to a commercial product of this kind would cover material of that kind, which is why we treat it as a prohibition rather than as a basis we rely on. The subject's agreement under clause 2.1 does not change that: agreeing to a report is not agreement to our holding material of these categories.

That prohibition is not yet enforced where it has to be, and we would rather say so than write the intended state in the present tense. Today there is no classifier on the write path. The check exists as a specification that would run after an assistant answer had already been persisted to the evidence store and to the model-call log, and our own engineering documents disagree about which stage owns it. We will not sell a report whose subject is a natural person until all four of the following are true and testable from outside:

1. a classifier runs on each retrieved item before anything is persisted;
2. an item it flags is refused at the write, so neither the database row nor the stored object is created;
3. the model-call log is written through the same refusal, so a prohibited answer does not survive there either; and
4. the refusal is recorded as a count of refusals by category, holding no copy of the refused text and nothing that identifies the subject.

Until all four hold, the Founder X-Ray does not open. When they hold, the release is recorded under clause 12.5, and until that entry exists a reader should take this clause as describing what we will do rather than what we do.

``

4. No human reviews a report before delivery

4.1 No person at LeMans Labs will read a report before it reaches the person who bought it. There is no editor, no reviewer, no sampled read and no sign-off. We state it plainly because a buyer paying for an assessment of a named person would otherwise reasonably assume somebody looked.

4.2 It is also why the rest of this notice exists. A system with no human in the loop has to be legible instead, which is why the rubric is published, the arithmetic is reproducible, and the failure modes in section 7 are written down rather than left to be discovered.

4.3 Every check before delivery is automated, and each is a gate rather than a warning. We publish what the checks test rather than how many there are, because the number will change as the pipeline changes and a count is the part of this clause a reader cannot verify. Each of the following will be tested on the finished draft: that every section required for that product is present; that every finding resolves to an evidence item the collection stage actually retrieved, on a page that was reachable when we read it; that every figure in the text matches a figure extracted from the evidence cited beside it; that the headline equals the plain mean of the six printed dimension values; that the source floor in clause 10.3 was met; that the report is not in substance an admission that the record was too thin to assess; that the subject was resolved from a strong identifier supplied with the brief rather than from a name alone, with the count of sources excluded as a different person printed in the report; and, for a report about a natural person, that nothing outside the permitted scope reached the document. The list in force is the list published here, and any change to it is recorded under clause 12.5.

What these checks are, and are not. They test a report against rules we can express in code. They do not establish that a statement on a retrieved page is true, and they do not establish that the subject resolved is the person the buyer had in mind, only that resolution ran against an identifier the buyer supplied and that ambiguous material was excluded and counted. Clause 7.1 explains why that distinction is the one that matters most.

What happens to a failed report, and to the money. A draft that fails a check is regenerated once. If it fails again the order is marked failed, the report is not delivered, and the full amount is refunded automatically, without a request and without a decision by anyone. We do not say a failed report is free: payment is taken when an order is placed, so what happens is that money leaves and comes back, and a refund reaching a card can take longer than the charge did. Section 2 of the Refund Policy governs the money and this notice does not restate it.

4.4 When a person at LeMans Labs does read a report. None of the following is review before delivery and we will not describe it as such. Each is a person reading what was written about a subject, so each is listed here rather than left to be discovered. A person reads a report to answer a correction from the buyer; to act on an objection, a rights request or a withdrawal from the person the report names; and to answer support. While a product carries a beta label, a person also reads delivered beta reports afterwards as a quality check, including reports about a named individual, and tells the buyer where we find something wrong rather than waiting for the buyer to find it; that read is set out in the Beta Features and Pre-release Notice and, because it means an employee reads the complete analysis of a named person, it is disclosed to that person in If a Report Names You as well as here. It ends when the beta label does. Where any of these reads changes what a report says, the report version increments under clause 12.1.

4.5 If we ever introduce a human read before delivery, the disclosure in clause 1.2 changes in the same release as the feature, on the terms in clause 1.5.

``

5. How to read the Intelligence Score

5.1 The rule. Six dimensions, each scored 0–100 against the rubric published at /intelligence-score. The headline Intelligence Score is the plain unweighted arithmetic mean of those six values, rounded once to the nearest whole number. No weighting, no curve, no adjustment of any kind. Every report will print all six values, so anyone holding the document can reproduce the headline with a calculator.

5.2 Where a dimension value comes from. Each dimension is five criteria drawn from the published "raises" and "lowers" text. Each criterion is judged on a five-point scale and converted to points by a fixed table:

Five criteria at up to 20 points each give the dimension its 0–100 value. We publish this table because the alternative is to call the rubric fully published while keeping the step that generates each number to ourselves.

5.3 What is reproducible, at its real scope. The arithmetic is reproducible: the six printed values always give the printed headline. The rule will be enforced in the database itself, so a report whose headline does not equal the mean of its six printed values cannot be stored as ready; that enforcement exists today as schema that no running system has yet applied, and it is one of the conditions of sale in clause 12.7. The judgement underneath is not reproducible. Two runs over the same frozen evidence with the same model will usually agree and may not; we have not measured how often they differ, and until we have we publish no figure for it. The seven samples on this site show the format and the rubric, and nothing else: their numbers were written by hand to demonstrate the format rather than produced by the pipeline, as clause 8.3 and Report Accuracy and Public Data explain, so they are not evidence about how the pipeline behaves.

5.4 Comparability, and its real limits. Scores produced under different rubrics are never comparable and will never be placed on a shared axis, chart or sorted list: a Company Intelligence Score of 79 and a Personal Brand Score of 79 are not the same statement about the world.

Within one rubric, two subjects are comparable only in the narrow sense that the same questions were asked of each and the same criterion scale was applied. The scores are not readings from a calibrated instrument. A difference of a few points between two subjects, or between one subject at two dates, may be run-to-run variation rather than a difference in the record, and we have not measured how large that variation is. Until we have, we state no threshold at which a difference becomes meaningful, we publish no figure for run-to-run agreement, and we sell no score movement as a product. Every report will print the rubric version, the pipeline version and the model that served each stage, so the conditions of any comparison can be checked rather than assumed.

5.5 What is printed beside the score. A number alone is not a measurement. Next to the headline, every report will print the number of items read, the number of distinct domains they came from, the date range of the material, the spread of evidence strength across findings, and the number of sources excluded because they could not be tied to the subject named in the order. Where the corpus is at the floor in clause 10.3 we will say so rather than showing the same clean two-digit number.

5.6 Bands. The four bands are defined by the rubric, never by a distribution of customers. We do not know and will never publish what share of subjects falls in any band, because that would be an invented number.

5.7 The Executive Package. The package is two reports, a Company X-Ray and a Founder X-Ray, and it prints twelve dimension values: the six of the company report and the six of the founder report. Each report carries its own Intelligence Score, computed under clause 5.1 as the plain unweighted mean of its own six printed values, and each of those is the score referred to in clause 5.1 and in the contractual minimum stated in the Refund Policy and in Report Accuracy and Public Data.

No third figure is computed. Until 14 August 2026 the package also printed a combined figure over all twelve values, and it has been withdrawn along with the analysis of the founder against the company, because neither of them could be produced without assuming that the person and the company named in one order are connected, and our ordering process does not establish that. The two scores are not compared with each other and their difference is not a reading of anything. Clause 5.4 governs both of them exactly as it governs any two scores from different rubrics.

6. What the score is not

6.1 Five statements, each of which narrows what we claim rather than what we owe:

1. Not a valuation. It measures the public record, not the business. A quiet company with excellent economics can score 58, and it should.
2. Not a prediction. Every dimension describes something already observable. Nothing in the score forecasts revenue, funding or outcome.
3. Not a ranking. We publish no leaderboards and sell no place on one. Comparison is the reader's own to make.
4. Not a judgement of quality. It reads how legible a subject is, not how good they are. The most common finding is a good business with an under-written public record.
5. Not a rating of a person, and no dimension scores standing. No dimension applied to a natural person scores that person's reputation, standing, character, conduct or worth. No criterion applied to a natural person scores them down because criticism of them exists in the public record, because coverage of them is unfavourable, or because something negative about them has not been displaced. Where a dimension applied to a person reads the public record about them, it measures one thing: whether that record describes their professional work consistently across independent source types. Criticism, disputes, complaints, adverse coverage, and the absence of a response to any of them, are not inputs, are not scored, and move no number in either direction.

6.2 A score about a person is a point-in-time reading of the public professional record on the date the material was collected. It is not a longitudinal behavioural score. We build no score histories, we do not compare a person's score at two dates unless a buyer orders a fresh report and we say so on its face, and we sell no product that tracks a person's score over time. Nothing accumulates from one report to the next: a second report about the same person reuses only the identifiers needed to resolve who they are, and stores no comparison, no history and no carried-over assessment. What we hold about a person a report names, and for how long, is set out class by class in the Data Retention notice rather than summarised here, because a period restated in two places drifts.

6.3 The score, the band, the summary sentence and the wording of every finding are output we authored. They are not a repetition of what somebody else said and we do not treat them as somebody else's responsibility. Where our own output about a person is wrong, correcting it is our job; the route is in If a Report Names You.

What kind of statement an assessment is. An assessment is our opinion, formed by applying a published rubric to material printed in the same document. It is offered as an opinion and not as an assertion of fact about the subject, and every assessment will be printed next to the evidence it was drawn from, so a reader can see what it rests on and reach a different view. Where a report records what somebody else published, we name the publisher in the sentence and we are stating that they published it, not that it is true. Report Accuracy and Public Data sets out the two kinds of statement and how each is corrected, and every finding in a delivered report will carry which of the two it is. That label, and the rule that the evidence prints beside the assessment, are conditions of sale under clause 12.7: until they ship, this paragraph describes what we will do.

6.4 Clause 6.1(5) is a condition of sale, not an intention. Until every dimension applied to a person in the rubric published at /intelligence-score satisfies clause 6.1(5), we do not sell the Founder X-Ray or the Executive Package, and no rubric version that fails it is applied to a natural person. Adverse characterisation of a named individual sits on the prohibited list in clause 3.7 and is refused at retrieval on the same terms, rather than collected and then left unscored. Where a criterion is removed for this reason, clause 12.1 records the rubric version at which it went.

``

7. Known failure modes

Named without hedging, because a list of failure modes that reads like marketing is worse than no list.

7.1 The wrong person. Two people share a name, sources about one are read as sources about the other, and a number verified against a page about the wrong person verifies perfectly. This is the most damaging thing the system can do and the least likely to be noticed downstream. We will resolve a person from a strong identifier rather than from a name, refuse a brief that supplies a name and nothing else, and print the count of sources excluded as a different person rather than absorbing them silently. The consent step in clause 2.1 is a second guard on the same failure, because the person who confirms is the person the report is bound to.

7.2 Faithful repetition of something wrong. Where a public source is wrong, a report can record the error accurately. What happens then is in Report Accuracy and Public Data and, if the report is about you, in If a Report Names You.

7.3 Evidence strength is not a probability of truth. Every finding carries a label we call evidence strength. We use no other word for it anywhere, because "confidence" invites a reader to hear a likelihood that the statement is true, and it is not one. It is computed in code from how many sources support the finding, how many independent registrable domains they come from, how recent they are and whether they disagree. It measures evidential redundancy: how much material says the same thing, not whether the material is right. Four outlets repeating one original error produce a high evidence strength on a false statement. The thresholds that turn the calculation into a label are a documented calibration and not a measured one; no error rate has been measured and we publish none. Read a high label as a reason to check the sources rather than as a reason not to.

7.4 A number that does not follow from the report. A dimension value comes from criterion judgements, and material can be removed from a report after it has influenced one. A number a reader cannot trace to something printed is a defect. Every dimension value will be required to rest on at least one judgement whose deciding evidence is cited by a finding that survived into the delivered report, and that evidence will be printed beside the dimension.

7.5 A thin record read as an absence. Not finding something is not evidence that it does not exist. Where the record is thin the report says what was searched and what was not found, and nothing else. We make no claim to have read everything, and a subject's absence from a source we read means only that.

7.6 Pages written to influence us. Someone can publish a page built to be read by our system and rank it for a subject's name without owning that subject's site. Our defences are aimed at provenance and arithmetic: retrieved text will never be treated as an instruction, hidden text will be stripped, the analysis stages will hold no tools, output will be structurally validated, and no model will produce the headline. None of that establishes that a statement on a real page at a real URL is true. In the other direction, a subject who reads the published rubric can build pages against it, and where we can see that pattern we will report it as an observation about the record rather than score it.

7.7 Variation between runs. The same subject analysed twice can produce slightly different dimension values, because the judgement underneath is not fully deterministic. We do not promise the same number twice, and we will not sell score movement as a product until we can say what a real movement is.

7.8 Time. A report describes the public record as it stood on the collection date, which will be printed on it, and begins going out of date from that moment.

7.9 Each of these has a route out, and the routes run in this order. Where a finding is wrong, tell us and we will re-examine the evidence bound to it: Report Accuracy and Public Data if you bought the report, If a Report Names You if it is about you. A correction is free, is not conditional on our agreeing with you, and is the fastest of these routes. Naming a failure mode in this section is not a statement that it occurred in your report and is not by itself a ground for a refund. The Refund Policy governs the money, including the automatic refund where a report fails the checks in clause 4.3 and where a consent request under clause 2.1 is not confirmed. Nothing in this section is a precondition to, or a substitute for, a consumer buyer's rights under the law, which are dealt with in clause 10.4.

8. How our output is marked

8.1 Report text is machine-generated and we will mark it as machine-generated in a form a machine can detect: metadata in the report page, and structured data identifying the document as an automated analysis and pointing back at this notice. A file exported from a report carries printed marks rather than a machine-readable one. Every page of it shows the report version, the time it was generated, a confidentiality mark, the sentence in clause 9.1 barring use for screening, and the address of the notice written for report subjects. We do not promise a metadata property inside the file that a machine can read, because we have not built one, and clause 8.2 is why we would not want such a sentence read as a watermark in any case.

8.2 Marking text is not watermarking and we will not imply otherwise. There is no reliable watermark for prose. What we mark is the container, so copying words out of a report leaves the marker behind, and a reader who receives report text second-hand cannot tell from the text alone where it came from.

8.3 The seven sample reports on this site were written by hand to demonstrate the format. Their dimension values, headline scores, source counts and corpus captions are synthetic, as the header of each sample states and as Report Accuracy and Public Data sets out. Because they are not output of the pipeline, they will not carry the marking in clause 8.1. Marking hand-written copy as an automated analysis of a named subject is a false statement in the other direction, and on a page naming a real person it would turn a demonstration into an apparent assessment of them. Each sample will instead carry, in the same machine-readable place, a marker identifying it as a synthetic demonstration of the report format and not an assessment of the named subject, so that a machine reading the page reaches the same conclusion as a person reading the header. Our marketing pages are written by people and carry neither marker, because they are neither. If we later publish samples the pipeline produced, those pages carry the clause 8.1 marking and this clause is reissued under clause 12.5.

8.4 We treat the machine-readable marking of content our system generates, and the disclosure of the automated nature of what we produce, as obligations that fall due when we place this system on the market. We build the test that proves we meet them before we sell, rather than treating either as deferred, and we rely on no transitional or grandfathering provision. We state no date in this notice for the application of any statutory obligation, and we cite no provision we have not had confirmed. We do not describe ourselves as compliant with any AI, data protection or consumer regulation, here or anywhere else. We describe what we do, and build the test that proves it.

``

9. Regulated decisions, and what the product must not be used for

9.1 This clause will be reproduced word for word wherever the restriction appears, including at checkout, on the report itself, on every page of a file exported from a report, and on every view served through a share link:

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.

9.2 A report must not be the sole or determining input into any decision about a person. It is one document about a public record, produced without human review, and it is not built to carry a decision on its own.

9.3 LeMans Labs takes no decision about a person a report describes. We decide nothing about a subject, grant nothing, refuse nothing and rank nothing; we apply no output to a subject; and we produce no threshold, flag or list that we act on. We assess a public record, produce a document, and deliver it to the person who asked for it. This is profiling and we call it that. We state those facts rather than the legal conclusion they are said to support, because the conclusion is not ours to certify and because a person a report names is entitled to test it against the facts rather than accept our characterisation of them. Nothing in this clause is a determination of anyone's rights, nothing in it is intended to tell a person a report names that they have no rights against us, and we will not argue that anything in this notice narrows them. Those rights, and how to exercise them without an account and without a fee, are in If a Report Names You.

The buyer's position is different, and it is theirs. A buyer who uses a report to decide something about a person is the decision-maker in that decision. The buyer may carry obligations towards that person that we do not, including obligations to explain the decision, and we supply none of the documentation, the accuracy standard or the process those obligations would require. If what a buyer needs is a document that can carry a decision about a named person, this is not that document, we do not sell it as one, and clause 9.1 is on the report and at checkout for that reason.

9.4 What the product is placed on the market for, and how that is kept true. This product is sold for company and market diligence: understanding how a company or a founder is represented in the public record, ahead of an investment, a partnership, a commercial engagement or a competitive review. It is not placed on the market for employment screening, recruitment, candidate assessment, creditworthiness assessment, insurance pricing or tenant screening.

We do not market it to recruiters, employers, human-resources functions, hiring or applicant-tracking platforms, lenders, insurers or landlords, and we name none of them as an audience or a use case on any page of this site, in any structured data we publish, in any advertising keyword or audience list, or in any invitation we send. A surface that names a barred audience is a defect in that surface. We correct it rather than defend it, and correcting it does not narrow the bar in clause 9.1, which does not depend on what any page of ours says. Tell us at [email protected] and we will remove it, record the correction with its date, and treat the removal as a material change under clause 12.5.

9.5 We will build no feature that turns a report into a decision: no threshold alerts, no pass or fail flags, no sorted lists of people, and no interface returning a score about a person without the report and this notice attached. We do not certify our own classification under any AI regulation and nothing here should be read as doing so.

9.6 The full list of subjects we refuse and uses we prohibit is in the Acceptable Use Policy, which governs. This section is the part of it that follows from how the system works.

9.7 Screening, and where we do not sell. Where we screen an order against a restricted-party list, that screening is a gate on whether the order proceeds and nothing else. It never becomes evidence, it is never scored and it never appears in a report, so the statement in Report Accuracy and Public Data that a report queries no criminal record, court record, credit file, licensing register, sanctions list or watch list stays true. The lists, the thresholds and the buyer's own obligations are in the Acceptable Use Policy, which governs. Separately, there is no territory into which we decline to sell. Where we sell is a commercial question carried by the Terms of Service and never a statement about what the system can do.

``

10. Responsible AI: the commitments we make, and the ones we do not

10.1 What we do not have. We hold no certification of any kind for this system. There has been no independent evaluation, no third-party audit, no external red-team exercise, no published fairness audit and no bias metric, because none has been measured. We carry no insurance for this activity. We do not describe ourselves as compliant with any data protection or AI regulation. Where this notice describes a practice, it is a practice and not an attestation.

10.2 What we commit to. Each is a rule with a mechanism behind it rather than a value statement:

1. The scoring rubric is published, including the criterion scale in clause 5.2.
2. The headline arithmetic is computed in code from the printed values and can be checked by anyone holding the document.
3. Every finding binds to the evidence it came from, and a finding that cannot be bound does not ship.
4. The absence of human review is disclosed rather than hidden, inside the disclosure itself.
5. For a person, collection is confined to the public professional record by a control at the point of retrieval, and no report about a natural person is sold before that control exists.
6. Reports are delivered privately, and stay private. We do not publish a report, index one, list one or sell one. Reports and share routes exist from this release and the search-engine work does not: robots entries, no-index directives, sitemap exclusion and a link preview carrying our name and nothing else are not built, and no test asserts any of them. What keeps a report out of an index today is that nothing anonymous can reach one. A shared report is served only to the person the link names, after that person has confirmed the address the link was sent to with a code, so a crawler, a preview fetcher or a platform unfurling a pasted address gets nothing. The outstanding work has its own row in section 13 of the Security notice and is stated in clause 12.6 of the Intellectual Property notice. We disclose a report to nobody but the buyer who ordered it, except to a processor listed in Sub-processors acting on our instructions in order to produce, deliver or store it; to the person the report is about, on their request; and where the law compels us, in which case we tell the buyer where we are permitted to. The buyer's onward disclosure is not at large, and the answer depends on who the report is about. A report about you is yours: you bought it, it is your own record, and you may show it to anyone you judge fit, investors and a board included. A report about a company is not. It goes to your own people who need it for the decision it was bought for, and to your own professional advisers under a duty of confidence, and no further. Either way the buyer may not republish it and must withdraw it on our written request. The Terms of Service sets those restrictions out and governs. From this release a buyer can export a report as a file and give a link to one person they name. The link is not anonymous: the buyer names the recipient and gives us the address, we send the link, and the recipient confirms that address with a code before the first view – the same mechanism a buyer signs in with. We serve every view ourselves, we record each one, a link never outlives the ninety-day window, and revoking one stops the next request. The file is the opposite: once it is exported we cannot recall it, revoke it or expire it, and the last sentence of this item is the whole of what we can honestly say about it. Where those limits are broken we end the licence, and we say plainly that ending a licence does not recover a file that has already left our systems.
7. Everyone who operates the pipeline will be briefed on how it works, how it fails and what it must not be used for, recorded with a date.
8. Before the first report about a named person is sold we will complete and sign an assessment of the impact of that product on the people it analyses, covering what is collected, why, the harm the product could do to them, and the measures that reduce it. A person a report names may ask us for a summary of that assessment, and we will send the summary within one month of the request; where we withhold any part of it we will say which part and why. Until the assessment is signed we sell no report whose subject is a natural person, and this item states a date only once the assessment exists.

10.3 What we do not promise, and the one quantity we do. We do not describe a report as accurate, complete, verified or free of error, and we do not use those words about our output, because the public record it is built from is none of those things. We do not promise the same score twice, for the reason in clause 7.7. We do not promise that a source we relied on is telling the truth. We publish no expected, typical or maximum number of sources.

One quantity is an exception, and it is a contractual minimum rather than an expectation. A Company X-Ray will read at least 40 distinct sources and a Founder X-Ray at least 25, counted as clause 12.3 counts them, and both floors apply to the Executive Package. That is the same floor as the one stated in the Refund Policy and in Report Accuracy and Public Data, and it is stated identically in all three. It is enforced as a gate rather than described as a target: a report that cannot reach it fails the checks in clause 4.3, is not delivered, and is refunded automatically. It is a floor and nothing else: not a typical figure, not an expected figure, not a target. Every report prints the number of items actually read and the number of distinct registrable domains behind them, so the floor can be checked against the document rather than against this page. Where any page of ours, past or present, states or implies a different figure, this clause governs, and a buyer who bought while a different figure was published may rely on whichever is more favourable to them.

10.4 This notice creates no exclusion and no limitation of anything we owe a buyer, and it is not a term anyone is asked to agree to in place of their rights. We do not rely on a sentence saying that a term applies only so far as the law allows: the terms of this notice are written to be fair as drafted, and if one of them is not, it falls and the rest stands.

If you are a consumer buyer: your right to receive a report matching the description you were given, your remedies if it does not, and your right to rely on the public statements we make about the product are set out in the Terms of Service and the Refund Policy. They come from the law rather than from us, they are not conditional on our agreeing with you, and we do not decide whether our own product conformed. Clause 10.3 states what our output claims to be; it disclaims no obligation and does not touch the description you were given before you bought.

If a report is about you and you bought nothing: you are party to no contract with us, you have agreed to nothing beyond the consent in clause 2.1, and no limit we agree with a buyer applies to you. Nothing in this notice, in our terms, or in any other document we publish for buyers affects your rights, and we will not argue otherwise. Those rights, and how to exercise them, are in If a Report Names You.

``

11. Providers and routing

11.1 We host no models. Every model call will go through a single routing provider, which forwards it to the company that serves the model, and search results will come from a third-party search API. The recipients, what each one receives, where it runs and the safeguard relied on for each transfer are set out in Sub-processors, which is the list that governs and which states, row by row, what is settled and what is not.

That list is not complete today, and you should hear that here rather than discover it there. No data processing agreement has been executed with any recipient. Several locations are stated as intended rather than configured. One component that would hold the complete rendered analysis of a named person has no location chosen at all. And the companies our router forwards to cannot be named until the permitted list is pinned by name in each request. Until every row a report about a natural person would touch is settled, and its safeguard actually in place, we do not sell such a report. Clause 12.5 records the release in which that becomes true.

11.2 Two consequences a report subject should understand plainly. To produce a report about a named person, that person's name leaves our systems: to a search provider as a query term, and to model providers as text inside the material we send them. And a page we fetch sees a request from us, so reading a subject's own site discloses that the site is being read. Both are stated in the consent request under clause 2.1, before anything is collected, so that a person deciding whether to confirm is deciding with those facts in front of them.

11.3 What the assistant probes measure, and what they are not. Some reports will record how AI assistants answer questions about a category and about the subject, and a report about a person will send prompts containing that person's name to each of the assistants named in Sub-processors.

What we record is the answer returned by the thing we queried, through the interface we queried it through, on the date we queried it. The report will print exactly that: what was queried, identified as precisely as the interface allows, and the date. We will not describe ourselves as having measured anything we did not query, and in particular we will not attribute an answer to an underlying model where what we queried was a consumer assistant product, or to a product where what we queried was a model interface: those products apply their own retrieval, ranking and instructions, and an answer is the output of all of that together. A probe result is therefore evidence of one answer, from one thing, at one moment, and of nothing else. It is not a measurement of a model, not a prediction of what the same thing will say next time, and not a measurement of the subject. Vendor and product names identify what was queried and imply no affiliation with, or endorsement by, their owners. We publish no comparison or ranking between them.

Two things about the probes are not settled in code today, and we would rather publish that than a tidy sentence. Whether these providers are reached through our routing provider or called directly is not fixed, and our own engineering documents disagree. Nor is the number of prompts sent for each product, and how many of them contain the subject's name, fixed between those documents. Both will be settled, published at /intelligence-score, and stated in Sub-processors before the first report about a natural person is sold, because a person who is asked to consent under clause 2.1 is entitled to be told how many times their name will be sent and to whom.

11.4 Training. Our routing configuration will be set to exclude providers that retain or train on the traffic we send, the permitted providers will be pinned by name, and an automated test will assert the pin. That is a commitment, not a present fact, and we will not claim that inputs are excluded from model training until the pin and the test exist.

11.5 What leaves our systems, stated once and governing. To produce a report we send the subject's name and the terms of the brief to a search provider as query text, and the material we retrieve, which contains the subject's name, to model providers as the text they read. Every recipient is listed in Sub-processors. We sell nothing a buyer submits, we train no model of our own on it, and we use it for nothing except producing and delivering the report and keeping the records described in the Data Retention notice. Any statement anywhere on this site that briefs, evidence or report material are never shared with third parties is wrong, and this clause supersedes it: material leaves our systems to the recipients named in the sub-processor list, and there is no configuration of this product in which it does not. Where you find such a statement, tell us at [email protected] and we correct the page.

``

11.6 How long the machinery keeps what it handled. A report, the material behind it and the files generated from it are deleted on the schedule in the Data Retention notice, which governs and which states each period once. Two consequences belong here rather than in a schedule. First, the log of the automated calls made while a report was produced is kept longer than the report itself, and a prompt inside it can contain a report subject's name; this is the section that tells a subject their name left our systems, so it is also the section that says the record of that call outlives the report. Second, once the material is deleted we cannot re-examine a finding, reproduce a source or explain how a statement was reached, and our copy is the only one we can reach. A file the buyer exported before deletion outlasts ours: it sits on their machine, outside anything on this page, and we cannot read it, correct it or take it back. So a person can be holding a finding we are no longer able to examine, and we say that rather than let this clause imply that our deletion ended the document. If you want a finding re-examined, tell us inside the retention window. After it we will tell you that we cannot, rather than answer from something we no longer hold.

``

12. Versioning, and how we count

12.1 Three version numbers will travel with every report and be printed on it. Rubric version changes when a dimension, its question, its criteria or its scoring text change. Pipeline version changes when the stages, prompts, models or retrieval behaviour change. Report version increments when a correction is issued for the same purchase.

12.2 Every report will also print its identifier, the date the evidence was collected, the date the report was generated, and what served each stage, including any substitute used when the first choice was unavailable. Where evidence collected earlier is reused, the date printed is the date that evidence was collected, because presenting old collection as fresh would be an invented measurement. For a report about a person, evidence will be collected fresh for each report and not reused.

12.3 How we count sources. A source is a distinct registrable domain, counted after collapsing syndication, so several outlets carrying the same wire copy count once, and a subject's own pages count as one source however many we read. Reports will print both figures: items read, and distinct domains. The floor in clause 10.3 is counted on the distinct-domain figure and not on the count of items read, so a report that read sixty items across thirty domains has not reached a floor of forty and is not delivered. What those counts do and do not tell you is in Report Accuracy and Public Data. We publish no expected or typical number of sources.

12.4 Reporting. We publish no statistics about reports produced, corrections made or objections received, because none exist yet, and a number invented for a transparency report is the same defect as a number invented for a marketing page. The method of counting is published here now; the first figures will follow the first full quarter of sales.

12.5 Changes to this notice. The version of this notice in force when a report was bought governs that report for its whole life. We will not change it for a purchase already made, and we will record which version was shown to a buyer against their order so this can be checked rather than asserted. A material change is published on this page with a new effective date, on the notice period in the Terms of Service, and does not apply retrospectively to a report already bought. A change to the disclosure wording in clause 1.2, to the scoring rule in clause 5.1, or to the position on human review in section 4 is always material. A change that softens any of the three elements fixed by clause 1.5 will not be made at all: if that position ever changes, it is a new notice under a new name and not a revision of this one. Every version is dated in this clause, and superseded text is available on request.

12.6 Questions. Write to [email protected] about anything to do with information we hold, and to [email protected] about anything else. Those are the only two addresses we publish.

``

12.7 The controls that are conditions of sale. The following controls are conditions of selling the product each applies to. None of them runs today. Each will be recorded in clause 12.5 with the release that met it, and we do not sell the product a control applies to until it is met.

1. The single disclosure string, the render gate over every surface in clause 1.2, and the test asserting it (clause 1.6).
2. The two pre-purchase placements of the disclosure, and the record of what a buyer was shown (clauses 1.2 and 1.3).
3. The automated eligibility and consent step, its confirmation window, its automatic cancellation and its automatic refund (clause 2.1).
4. The prohibited-category control bound to the evidence write, in all four limbs (clauses 2.4 and 3.7).
5. The release checks, and the database enforcement of the headline rule (clauses 4.3 and 5.3).
6. A rubric in which no dimension applied to a natural person scores standing, and the claim-type label and evidence-adjacent rendering described in clause 6.3 (clauses 6.1(5), 6.3 and 6.4).
7. The marking of generated output, the synthetic-demonstration marker on the samples, and the contract test asserting both (clauses 8.1 to 8.3).
8. The signed impact assessment for products that analyse a named person (clause 10.2, item 8).
9. The indexing exclusion over report, share and file routes, and the test asserting it (clause 10.2, item 6).
10. The pinned provider allow-list and the automated test asserting the pin, and every sub-processor row a report about a person would touch settled with its safeguard in place (clauses 11.1 and 11.4).

---

This notice is part of the LeMans Labs legal set and should be read together with Report Accuracy and Public Data, which covers what a finding is and what happens when one is wrong; the Acceptable Use Policy, which governs who a report may be about and what it may be used for; If a Report Names You, written for a person who did not buy anything; Sub-processors, which names every recipient; Data Retention, which states how long each class of material is kept; the Privacy Policy; the Refund Policy; and the Terms of Service, which is the contract and which also carries our provider identification and the address for notices.