RENDA standard for verifying AI-assisted documents RENDA-STD-VERIF-01 · Edition 1.2 · English Nature of this document. This public reference describes the issuer’s verification method. Readers, including those using AI assistance, may use it to reproduce or challenge a finding. They retain control over their tools, instructions and conclusions. Its evidentiary value depends on the sources and the checks actually performed for the document in hand. 01 | European legal framework and scope Regulation (EU) 2024/1689 establishes harmonised rules on artificial intelligence. Article 1 sets out its purpose, including support for human-centric, trustworthy AI and protection of fundamental rights. Article 4 deals with AI literacy in the context in which systems are used. For this editorial practice, the reviewer needs to recognise a fabricated citation, distinguish a quotation from an inference and know when to ask for the source. Article 5(1)(a) addresses prohibited subliminal or purposefully manipulative or deceptive techniques under specified conditions. These include appreciable impairment of informed decision-making and actual or reasonably likely significant harm. A disputed interpretation in correspondence does not by itself establish those conditions. RENDA’s own rule is to record and correct an omission that would materially mislead a reader. Articles 10, 12, 13, 14 and 15 set requirements for high-risk AI systems within their respective scope. They concern data and data governance, record-keeping, transparency and information for deployers, human oversight, and accuracy, robustness and cybersecurity. Article 14(4)(b) expressly addresses the tendency to rely too heavily on AI output. A RENDA reviewer checks the underlying record even when several AI-generated summaries agree. That is an editorial choice that can also be applied to ordinary correspondence. Annex III, point 8(a), concerns specified uses by or on behalf of a judicial authority and similar use in alternative dispute resolution. A party’s use of drafting assistance for its own submission is a different intended use. Classification and legal duties depend on the actual system, purpose, role and applicable provisions, including Article 6 where relevant. Article 50(4) addresses, among other things, AI-generated or manipulated text published to inform the public on matters of public interest. It provides an exception where human review or editorial control has taken place and a natural or legal person holds editorial responsibility for publication. RENDA also discloses AI assistance in letters to named recipients as its own practice. For a legal assertion in a case, cite the article, paragraph, applicable version and relevant date. The case evidence must separately establish whether its conditions were met. 02 | Fourteen codes used in this edition The following entries explain how this edition applies fourteen of the 119 codes in Virgil Profeanu’s Seven SoPh[A]iloTechnology Manifestos (2025). These are editorial applications. The operational view links to the fourteen English entries used here. The website also carries the full English text of all 119 codes. Zenodo holds the complete trilingual book. A case-specific factual assertion still needs its own exhibit. [FC 07] Match certainty to support. State whether the evidence shows a fact, supports an inference, permits an interpretation or leaves the matter unresolved. If a missing annex makes a condition uncertain, preserve that uncertainty in the conclusion. [FC 09] Keep a record that another reader can inspect. Carry the exhibit reference and its version from the working file into the audit record. Preserve the link to the version actually used when later editions appear. [FC 15] Test the ethical result of the drafting decision. Ask whether a reader would be led to believe something the file does not establish. Apply this test even when the more favourable account would be easier to defend. [PA 05] Check the completed document as a whole. Compare the subject line, summary, conclusion and attachments before release. [PA 07] Verify the smallest unit that carries meaning. A single digit in a date, the word “proposed” beside a price, or the difference between calendar and working days may change the practical effect of the sentence. [PA 08] Make completion a recorded decision. A draft passes the check when the named evidence supports the issued wording within its stated scope. Record a failed check and the action that resolved it. [PA 15] Give each ethical commitment an observable test. For an undertaking to quote accurately, another reader should be able to locate the passage, compare the words and see how material conditions were preserved. [PA 16] Leave a usable record for the next reader. Include enough context for someone coming to the case later to repeat the review without the original drafter, private notes or a particular AI service. [tCR 05] Have the issuer define the purpose of the communication. Record whether the document reports facts, advances an argument or answers a limited question. That purpose gives the reader the proper frame for judging completeness. [tCR 08] Preserve human ethical judgment at release. When expediency would favour a misleading deletion, identify the problem to the person responsible for the text and preserve the materially accurate account. [VC 12] Check the integrity of inputs and examine the output critically. Apparent agreement between two summaries may arise because both inherited the same transcription error. Return to the underlying record. [VC 13] Use assistance to extend the reviewer's capacity. A model can find dates to compare or suggest a counter-reading. The reviewer examines that evidence and records the judgment actually reached. [CL 13] Make the issued version identifiable. A recorded digital fingerprint enables a reader to compare a file with a reference copy. Record which file was measured and where a trusted reference can be obtained. [CL 16] Make a decision inspectable. A reader should be able to follow the data used, the rule applied, the material constraint and the resulting claim. Where the reader has limited access, identify the restriction and how authorised inspection works. 03 | Identify the claim, evidence and qualification Give each material claim a stable identifier. Record the source, date, revision and a usable locator, such as the page and paragraph, drawing detail or table cell and unit. Check an OCR quotation against the image. Mark a translation, illegible digit or missing page if it could change the finding. A documented fact has a traceable source. An inference shows how the evidence supports a further conclusion. An interpretation gives the technical or legal reasoning. An unresolved matter identifies the missing evidence. These four descriptions are editorial applications; FC 07 in the manifesto uses its own scale. A letter recording that a witness alleged a late delivery establishes that the allegation was made. Establishing the delivery time may require a receipt or timestamp. Where someone asks to shorten a statement, test what a reasonable reader would understand from the words that remain. If removing a condition changes the meaning, decline the misleading wording and offer a faithful alternative. A reply limited to the date of receipt may omit an unrelated issue when the limited purpose is clear. Include both the confirmation and the reservation when giving the recipient's complete position. Trace conflicting records to their date and subject and explain any difference still unresolved. 04 | Three checks before issue First check · Source. Locate each decisive exhibit or legal provision, confirm its identity and read the passage in the version cited. Compare the final quotation, number or date with the original. Record the locator and the outcome. If a material source cannot be inspected, identify the claim as unverified and either narrow it to what can be established or leave the issue open. Second check · Scope. Trace each claim from its source into the draft. Look for qualifications, disputed points, later corrections and evidence against the proposed wording. Check quantities against their units and population, deadlines against their triggering event, and duties against the party to whom they apply. Record the qualification that survives into the final text, or explain why a particular item falls outside the declared scope. Third check · Recipient reading. Give a reviewer the evidence the recipient is entitled to use, the stated purpose and the exact text intended for issue. For a blind review, withhold the drafter’s preferred conclusion, prior answers and private explanations. Ask the reviewer to record what an ordinary reader would understand, which passage supports that reading and what could change it. Compare the result with the evidence. An explanation in an internal chat cannot repair an omission in the document being sent. Record who or what performed each check and whether a person reviewed the final version. If a second AI pass used the same model, identify it as such. A fresh session can reduce exposure to the first draft's rationale while still sharing the first model's limitations. Describe independence only to the extent actually achieved. For material disagreement, preserve both findings, cite the decisive exhibit and record the human decision. Recheck affected claims after changing the final text. Record any unresolved point and its effect on release. 05 | Atomic unit of evidence and independent audit checklist For each decisive assertion the issuer records one evidence unit. Keep it alongside the final wording or identify how an authorised reader can inspect it. Claim ID and exact final wording | Fact, inference, interpretation or unresolved matter Exhibit and version | Page, paragraph or equivalent locator | Original words or measured value and unit What the evidence establishes | Material condition, counter-evidence or uncertainty Finding and responsible reviewer | Issued revision | Correction or access route Worked example. Claim A1: “The recipient confirmed receipt of the drawings.” The first sentence of the fictional exhibit, paragraph 1, supports receipt. The second records rejection of the proposed deadline. In a summary of the response to the proposal, the finding must carry both points. The published example card provides the two files separately so the reader can compare them. An independent reviewer can record PASS, REVISE or UNVERIFIED for each check. The following five questions apply across the three stages in section 04. Record a reason and locator for each material discrepancy. CHECK 1 · Source identity. Can I retrieve the cited edition and find the exact evidence? CHECK 2 · Accuracy. Do the quoted words, dates, amounts and attribution match the original? CHECK 3 · Scope. Does the text preserve conditions and material evidence supporting a different conclusion? CHECK 4 · Status. Can I distinguish direct evidence, inference, legal argument and remaining uncertainty? CHECK 5 · Issued text. Reading the final version on its own, would I infer more than the file supports? A blind reviewer receives the exhibits, the question being answered and the exact outgoing text. Record the reviewer's access limits. If a decisive exhibit is missing, mark the affected claim UNVERIFIED and identify who can supply it. A report mentioning a check is evidence of the review only when the underlying finding can be inspected. 06 | Reading a document with AI assistance The standard is available as plain text for direct reading, indexing and quotation. A reviewer may give the whole text to an AI assistant alongside the document and the exhibits that the reviewer is entitled to share. A review is reproducible when another reader can find the same passage and identify the same condition, even if that reader reaches a different legal conclusion. A recipient can supply the UTF-8 standard and the exhibits they may lawfully share to an assistant of their choice. A useful review identifies the passage relied on, evidence pointing the other way, and what remains unresolved. A disagreement with RENDA is welcome when the reader can trace it to the evidence. The public index lists the materials and the languages in which they are available. 07 | Audit record, access and corrections The record identifies the issuer, issued version, issue date and edition of the standard. For each decisive claim it gives the evidence unit, what was checked, who performed each check and which material questions remain open. Give readers a contact route for reporting a discrepancy against a specified claim and revision. The public card may link to a restricted record. Review disclosure rights before publishing names, excerpts or exhibits. Use effective access controls for confidential material. Record what a public reader can verify and how an authorised recipient can inspect the decisive source. If a correction is warranted, retain the issued version and its evidence. Link the dated correction in both directions and identify the claim and reason for change. 08 | File integrity and direct access The short URL and QR code open the audit card for the issued revision. Reserve that URL before inserting it into the document. Measure the exact final file after signature and record its SHA-256 in the card. A match to a trusted fingerprint supports byte identity. Issuer identity, approval time and substantive accuracy each require separate evidence. The English standard and all 119 English code entries are available as UTF-8 text and in a downloadable SQLite index. The full trilingual collection of 119 codes is available on Zenodo. Retrieve the whole language passage, including the generative principle where present. Cite the identifier, language and anchor so another reader can repeat the check. Primary references Official Journal · Regulation (EU) 2024/1689, original publication · English https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689 Official Journal · Regulation (EU) 2024/1689, original publication · Romanian https://eur-lex.europa.eu/legal-content/RO/TXT/HTML/?uri=CELEX:32024R1689 European Commission · Article 4 · AI literacy https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4 European Commission · Article 5 · Prohibited AI practices https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-5 European Commission · Article 14 · Human oversight https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14 European Commission · Article 15 · Accuracy, robustness and cybersecurity https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-15 European Commission · Article 50 · Transparency obligations https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50 European Commission · Annex III · High-risk uses https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3 Virgil Profeanu · The Seven SoPh[A]iloTechnology Manifestos (2025) https://doi.org/10.5281/zenodo.17012491