AI transparency notice

ECGT Ready uses an artificial intelligence component to classify environmental claims. This page is the transparency notice required by Article 50 of the EU AI Act (Regulation (EU) 2024/1689), whose transparency obligations apply from August 2026.

1. AI system in scope

• Component: claim detection, classification, and rewrite suggestion in the platform; per-page diagnosis and drafted fix paragraphs across privacy policy, cookies, terms, returns, contact, and product claim pages.
• Provider of the underlying model: Anthropic, PBC. A full audit runs in four steps, a first read, a verifier, two checks across the whole site and, since 21 September 2026, a final review that reads the pages of a paid audit a second time, each on a model of the Claude family (see section 1a).
• Deployer: ECGT Ready, operated by Mathys Seynaeve from Mamer, Luxembourg.
• Risk classification: limited risk under the AI Act. The system is not a high-risk AI system within the meaning of Annex III. It is not used for credit scoring, employment, education, law enforcement, or biometric identification.

1a. The steps of an audit

• First read: reads each page, detects the claims and the other risks in scope, scores severity and drafts the suggested rewrite; it also builds the claim register of the audit. Used on every plan.
• Verifier: checks each finding of a full audit against the allowed sources and attaches the citation. Used on every full audit.
• Checks across the site. One request compares what the pages say with each other and lists the statements that conflict. Another lists the elements the law requires that no page shows. Both read the start of each page. Used on every full audit.
• Final review: the most capable Claude model available to us at the time reads the pages again with the first read findings attached, confirms, dismisses, regrades or adds findings, merges repeats and writes the executive summary. Findings it removes are kept in an audit trail. Used on every full audit and zone audit.
• Guard check. One request reads a product page that Guard watches, when the page is new or its text has changed.

Each step runs on the latest version of the Claude model it uses, checked at every release of the platform; since 1 October 2026 the first read and the verifier run on newer versions than before. The models of this generation reason before they answer, within a bound we set for each step; that reasoning is part of the same request and is not kept. If the provider or the role of a step changes we update this page.

Since 30 September 2026 the pipeline reads the lines shared by the pages of a site (header, footer, cookie strip, banners) once, as one site wide page, instead of once per page; since 1 October 2026 the findings on them are reported under their own label, site wide elements. The navigation menus inside the header and the footer are the exception. They are left out of what the models read, and only the rules that run without a model read their labels. The fixed instructions of a step are cached by the provider for a few minutes so that the pages of one audit do not each carry a fresh copy of them; since 1 October 2026 one page is sent first and the others once its answer has started, so that they use that cache. The cache holds the instructions and what the pages of one audit share (the context of the site, the facts read across the crawl, your policy library when the audit uses it), never the text of a page. It is part of the same API processing as the request itself and is never used for training.

Since 1 October 2026 a page is not sent to the model twice for the same answer. When a page, the facts read from it, the rules, the model and its reasoning bound are exactly those of a read made for the same account in the last 90 days, the platform reuses the answer of that read (section 7). Any change to the page, its facts or the rules makes a new read. The final review of a paid audit always reads every page.

1b. What the platform reads on a page

Since 2 October 2026 the models get the text a visitor reads on a page. That is every heading, paragraph, list, table, label and button, the page title, the description the page gives to search engines, the text alternative of images, tick boxes and options with their state (ticked by default or not), drop down lists of up to 30 options, and content hidden from screen readers but shown on screen. Text that is in the page and hidden by its style, such as a closed tab or a collapsed answer, is read like the rest. Scripts, styles and embedded frames are left out, and so are templates no script has used, single characters, bare counts and the navigation menus inside the header and the footer (section 1a).

A page gives 600 lines at most, and a long page is read in part. The first read and the final review take the first 24,000 characters of a page and a Guard check the first 12,000. The comparison of the pages takes the first 9,000 characters of a page and the list of missing elements the first 3,500. Since 5 October 2026 both take the legal pages (terms, privacy, returns, cookies, contact) and the site wide elements up to 24,000 characters each, until this has added 80,000 characters to the check. A text that was cut ends with a mark, and each step is instructed to report nothing as missing on a page whose end it did not read.

The models read text. They do not see images, so a claim or a notice that exists only inside an image is read through its text alternative or not at all. Section 6 lists what else is not read.

On a full audit the platform also opens up to 30 pages in a browser, without a model. It keeps a picture of the top of each page, the names of the cookies present once the page has loaded and the other hosts the page contacts. It clicks nothing, so it makes no choice in a cookie banner.

2. Purpose

Detect environmental claims on a webpage, classify them against the rule set described in our methodology, and propose lower-risk rewrites. The output is an indicative risk score and a list of flagged sentences.

Since 1 October 2026 a workspace can add its own policy library (claims it forbids, wording it approved, certifications it may cite). When an audit or a Guard check runs with that scope, the same models also flag page text that breaks a line of the library and cite the line. Such a finding is the customer's own rule applied to its pages, not a statement about EU law, and it is labelled as such in the platform and in the reports.

Since 30 September 2026 the same models also check two further rules on every full audit: the online withdrawal function (Directive 2011/83/EU, Article 11a and Article 6(1)(h), applicable since 19 June 2026) and the prior price behind every announced reduction (Directive 98/6/EC, Article 6a, read with the judgment of the Court of Justice in C-330/23), and they treat a reference to the EU online dispute resolution platform as outdated since its closure on 20 July 2025. For these rules the platform first reads the page itself, without a model: the price blocks and the percentages recomputed on each price, the labels of links and buttons, the fields and the button of a withdrawal form, the pages the crawl covered. These facts and, where available, the prices the platform recorded on earlier reads of the same product page (kept 90 days) are given to the model next to the page text, with the rule and the fix templates. The model decides and drafts; the facts are what it decides on. No model, provider or model use was added.

Since 1 October 2026 each finding also shows what the law says: the model names the provision the finding rests on, and the text of that provision is taken from the platform's library of the official texts of the Publications Office of the European Union, quoted word for word. The model never writes the text of the law shown with a finding.

Since 3 October 2026 a few rules are applied without a model, on what the model returned. A claim that names a target year or speaks of the future is placed under the rule on future environmental performance (Directive 2005/29/EC, Article 6(2)(d)), whatever provision the model named, and each claim of the claim register gets the provision of its kind. The models also check, on the page that carries the legal guarantee, whether its text or the text alternative of an image shows the harmonised notice that applies since 27 September 2026 (Directive 2011/83/EU, Article 22a).

Since 13 October 2026 a paid audit of a storefront the pages confirm as French (a French company form, VAT number or address, never the language alone) gives the models the French rules next to the EU families, with the facts of the crawl they rest on (the price suffixes, the index pictogram and score, the words of an environmental claim, the guarantee box of the terms, the items of the mentions légales, the termination function). The same models do the same steps; no model, provider or model use was added. The French texts shown with a finding come from the platform's national library (the Legifrance data, read on a mirror on 13 October 2026), never from the model.

Since 13 October 2026 a paid audit of a shop whose pages sell a food, a food supplement or a cosmetic also looks each health claim up in the EU Register of nutrition and health claims and the EFSA list of botanical claims on hold, and checks each nutrition claim against the conditions of the Annex of Regulation (EC) No 1924/2006 with the figures of the page. The lookup runs without a model, on the words of the page and the register as data (exported from the European Commission's open data on 13 October 2026). Its outcome, the register entry and the condition met or failed, is given to the models as a fact next to the page text. One judgment stays with the model, whether a wording that differs from the authorised one has the same meaning for the average consumer, and the finding then says so. The same models do the same steps; no model, provider or model use was added.

Since 13 October 2026 the report of a paid audit of the pages of a shop ends with the label guide, three lists of labels built without a model from the words of the pages, the claims the first read listed and a register of labels kept as data (lib/labels/register.ts, 50 rows, the fees and the date each was read). No model call is added by the guide, and the models of the audit are not given the register.

Since 13 October 2026 a paid audit gives the models four more sets of rules when the pages call for them, the legal guarantee and digital content, geo-blocking, consumer credit and pay later, and the Data Act on connected products, with the facts of the crawl they rest on (the guarantee, residence, payment, pay later and connected product sentences of the pages). The same models do the same steps; no model, provider or model use was added. The texts shown with a finding of these families come from the platform's library of the official texts, never from the model.

2a. Reggie, the chat of the platform (27 October 2026)

Reggie is an AI assistant that answers questions on EU law and on the law of the countries listed on our libraries page, and checks the material you give it (a page by link, a product text, a campaign, an email, a label, the artwork of a packaging, a picture or a file) before it goes live. It is an AI system of limited risk under Article 50 of the AI Act, and the window says so once, in the same words as our FAQ, the moment you open it, so you always know you are talking to a machine and not to a lawyer. Every answer names the directive or regulation and the article it rests on, shows the date that text was read in the official legal library and links to the article at its source, and a second pass checks that the article as read supports what the answer says before the answer is shown. Reggie gives no legal opinion, and the model provider named in section 1 runs the inference without training on what you send.

3. What this AI is not

• It is not an automated legal decision under GDPR Article 22.
• It does not block your store, contact regulators, or fine you.
• It does not produce a binding compliance certificate.
• It does not generate deepfake media or impersonations.

4. Training and customer data

Anthropic does not train its production models on data submitted through the API. We do not run a separate training pipeline that feeds on customer data. If we ever introduce a feedback loop that adapts model behaviour, we will publish that fact here, ask for explicit consent, and offer an opt-out.

5. Human oversight

You decide what happens with every finding. Each one can be dismissed as not applicable, with a note, and reopened later. A dismissed finding leaves the open list and stays in the record of the audit. To challenge an output, write to contact@ecgtready.eu. For high stakes claims (future commitments, named certifications) we recommend a human review by qualified counsel before publishing. The disclaimer is shown next to every report.

6. Limits and risks

• The model can hallucinate. Flags should be checked against the source page.
• The model is trained on text up to a fixed cutoff. Recent regulatory guidance may not be reflected immediately.
• Pages written in several languages can be misclassified.
• Text inside an image, a video or an embedded frame is not read. An image is read through its text alternative, when the page gives one.
• Text a script loads only after a click, a scroll or a delay (reviews loaded on scroll, a banner shown after some seconds) is not read. Nor is the end of a page longer than the limits of section 1b. A page that shows no text without its script is opened in a browser and read as it renders, and listed as not read when that fails; text a script adds to a page that already shows text is not read.
• The platform reads public pages only. A withdrawal function or a price shown only after login is not read; the report then says so and lowers the severity rather than claiming an absence. The steps of the checkout are not read.
• An audit follows the links of the page it starts on and of the pages it finds. A page no link leads to is not read, and a page found beyond the size of the audit is listed as found and not read.
• The browser records the names of cookies and the hosts a page contacts. It does not know what a cookie is for or what a request carries.
• A price history exists only from the day the platform first read a product page for you. Before that, a struck price is reported as unverified, never as a proven breach.

7. Data minimisation

We send the model only the page text and minimal metadata required to classify (URL, locale, category), the facts read from the page for the prior price and withdrawal function rules (prices, labels, form fields, crawl coverage) and the prices recorded on earlier reads of the same product page, plus the lines of the workspace policy library when the audit runs with that scope. Where your workspace holds them, the profile of the business (country, industry, size, sales channels, markets), the packaging profile and the labels and descriptions of your supporting documents are sent too, never the files. We do not send account identifiers or billing data. For each call we keep a metering record (model, step, reasoning bound, page URL, token counts, cost and, since 2 October 2026, whether the answer could be used as it came or had to be cut, in counts, never in page text), never the prompt. Since 1 October 2026 the answer of a read (the findings of a page or the verdict on a finding) is kept for 90 days with a fingerprint of the request, never the request itself, so that the same read is not made and paid twice.

8. Evaluation

We test the platform on storefronts we wrote for that purpose. Each planted issue is listed in an answer key, next to a twin storefront that carries none. We run them when a rule or a model changes and count what was found, what was missed and what was raised wrongly. The results are used to adjust the rule set.

9. Contact

Questions, feedback, or to challenge an output: contact@ecgtready.eu. For privacy-specific concerns: contact@ecgtready.eu.