Preparation, query and validation

How IUCOM Reporter turns facts into a supported answer – and how it checks every answer against the sources themselves

Version 1.3 · 9 October 2026 · Process dossier

1The path of a request at a glance

Everything that leaves the computer is listed in Table 1a – always with your own key or directly to the official body or the catalogue and DOI service, without a detour via deepnet. The mapping of the placeholders never leaves the computer.

2Part I – Preparation: protecting data before sending

2.1Placeholders in four layers

Before anything leaves the computer, the program replaces personal and identifying data with placeholders such as [PERSON_1], [FIRMA_2] or [IBAN_1]. The replacement is reversible: the mapping stays local and is never sent.

Not replaced are citations of provisions (§ 1295 Abs 1 ABGB), amounts, years after “since”, “until” or month names, and job titles – otherwise the facts would lose their legal content.

2.2Check before sending

If anything was replaced or suspected – or “always show the check” is switched on –, the program shows the version that goes out: in a separate Word instance (if LibreOffice is set or Word is missing, in the program window) with the placeholders in red, beside it a table with tick, item in the text, placeholder sent, type and how it was detected. Uncertain items are yellow at the top (“unchanged – please check”). Further names can be added in one line. Nothing is sent before “Send”; “Cancel” ends the request without anything going out. Cautious users switch on “always show the check” in the options: then every version appears before sending, even if nothing was recognised, and missed names can still be replaced.

2.3AI preparation: the facts become the search

The chosen AI reads the checked placeholder version and returns in fixed lines what the search needs. It also decides which source is asked at all: without a tax link the Findok stays off, without an EU-law link EUR-Lex.

The instruction to the AI is written with real umlauts – with ASCII examples the AI wrote “ae/ss” and the RIS found nothing. Very short inputs (under five words) are expressly expanded. If the AI returns nothing – no key, empty answer, wrong form –, the search falls back to the words of the text, and a diagnosis reports the reason; it also checks whether the RIS is reachable and whether filters or too many words emptied the search.

2.4Area of law, rules and own documents

One or more areas of law can be chosen; their additional rules and documents go along, the choice itself only as a pointer – the facts are decisive. Own documents (firm standard, leaflets) are searched in two ways: by wording (BM25) and by meaning – a small language model (multilingual-e5-small) runs locally; the mix is 60 % wording, 40 % meaning. Observed on one real document (small sample): paraphrased questions found the passage in 3 of 5 instead of 1 of 5 cases. Only the matching excerpt with its heading goes along. The documents of a chosen area of law go along in full as long as together they do not exceed 24,000 characters; otherwise they are split into sections of about 1,500 characters and those best matching the facts go along (rare terms count more), in document order. Previously they were cut off in sequence – what stood at the back of a large document never arrived. Tested with test documents: the key passage at the end of a large first document, in the second and in the short third document went along each time (4 of 4 cases).

Rights before every import. Before documents are added – documents, gold standard and the works of a style profile when it is created –, the window “Confirm the rights to the documents” appears: the list of files with a traffic light per file from the local licence check running in the background (one reading, not legal advice), and the mandatory tick “I confirm that all rights to these documents have been cleared beforehand and that I am liable for their use myself”. For documents marked red – classification “permission required”, e.g. publishers' works, journal articles or excerpts from licensed databases – there is an additional tick “a licence or permission of the rights holder exists”; without it they are added but not sent to the AI, not even as gold standard or into a style profile. Every confirmation is recorded with time, Windows user, purpose, file names and classification in the program folder under Protokoll\rechtebestaetigungen.log; there is no second question when the files are read in. The manufacturer accepts no liability for the rights to imported texts (licence agreement 1.2, clause 4.4).

3Part II – Query: sources, relevance and request

3.1The sources and when they are asked

Each source returns three times what goes out; what counts is how well an entry fits the text – a match in the provisions or the case number counts double, the same work from several services once. The RIS is queried with throttling (at most 0.5 requests per second, no parallel connections) – this applies to every RIS call, including the case-law search per court; the other sources run alongside at the same time. A short network outage is retried once after 1.5 s, but not after “Cancel”.

3.2Relevance check by the AI

Before the statutory texts are added, the AI checks every entry against the facts and discards what does not fit; chapter 1 of the report names what was discarded. This happens before the statute search so that the provisions of discarded entries are not added at all – in a medical liability case, for instance, no asylum provisions entered the request any more.

3.3The request to the AI

The request consists of the rule set (expert-opinion style, interpretation by rank and method, scope as chosen), the additional rules of the chosen areas of law, own documents, the block of entries and the facts, plus a checklist of sub-questions. Every entry stands in the block with origin, court or publisher, number, date, link and licence notice – with the rule to cite only from it; literature is an aid to interpretation and is cited with author, year and DOI only if it is among the entries. The last rule requires a plausibility test of every word, also in the facts: what is unclear is reported rather than guessed.

Style adoption (options → input/output tab, off by default): with the tick “Write the answer in the style of a style profile” the AI writes in the style of a particular person. The button “Style profiles …” opens a window with the profiles per person – one folder Stil\<name> each in the program folder – and the steps create, open folder, create style profile, use for answers and switch off. To create a profile the program selects excerpts from the works locally (beginning, end and evenly spread paragraphs; at most 40,000 characters, 2,500 to 12,000 per work, at most 40 works, without file names) and measures figures locally: sentence length, nominal style, passive voice, perspective, connectives, spellings such as zB, iSd and Rz, quotation marks, structure. Then follow the confirmation of rights (2.4), pseudonymisation as for the documents and – always – the check before sending; then it goes once to the AI provider set. The provider writes a style profile in its own words – tone, sentence structure, choice of words, structure, argument, citation style, what the person avoids, plus a writing instruction –, without names, without words or phrases from the samples verbatim, at most 900 words; the costs appear in the cost book under “Stilprofil”. Afterwards only the profile goes along with each request, never the works and never the profile's name. Content, binding to the sources, citation and checking rules and the prescribed form take precedence; in correction mode the profile does not apply. Only own works or works whose author has consented may be used.

4Part III – Validation: checking against the sources

4.1The yardstick

AI answers in law rarely fail because they miss the one right solution – often several are tenable. The danger is that they assert something false: a decision that does not exist; a real case number with a wrong date; an entry that does not support the statement; a provision from the wrong legal order or in the wrong version; a fact that is not in the facts. All of this can be checked against the sources themselves – without a model solution.

4.2The checking rules in the program

Not supported is not the same as invented: only what demonstrably does not exist counts as invented. An assessment without supported entries carries a notice at the top; the source chapter and the attribution of all sources used are appended by the program itself, not by the AI.

Since 7 October 2026 literature appears per work in the verification report, since 10 October 2026 also from the running text; the summary names it in one sentence: “Literature: N references – x from the sources or your documents, y NOT AMONG THE SOURCES (added by the AI), of which z not found in the catalogue or faulty.” A reference that is in none of the sources provided comes from the AI's training knowledge – this is exactly where invented works and misattributed margin numbers arise; it is therefore reported even if the work exists. Whether the paragraph or page supports the statement remains open – the content of commentaries is never read.

The existence check of provisions queries the version in force: a provision that applied in the assessment period and has since been repealed appears as “does not exist”, as does a section of a foreign legal order with the same abbreviation (e.g. § 32d EStG of German law) – the status therefore asks expressly whether another law is meant. Laws outside Austrian federal law (treaties, EU law) are not checked by this rule.

4.3The verification report

The result is available three times: at the end of the answer document (Word or LibreOffice Writer) as the table “Verification report of the sources” with no., citation, footnote with page, status in its colour and the full source with link; in the archive as chapter 6; and as a separate part “Verification report” in the archive window (HTML, Word, PDF). Citations without support are red in the answer text. The AI's footnotes become real footnotes on the page where they are referenced; no paragraph is split across a page break.

4.4Checked where everyone has the same basis

Case law and statutes are the common foundation on which every citation must ultimately rest – there the Reporter is fully connected (RIS, the German federal courts and statutes, EUR-Lex, CJEU and ECtHR). For legal literature the gap remains that everyone has: commentaries, handbooks and articles sit in publishers' databases. The Reporter names this openly instead of hiding it, and narrows it: the literature check verifies whether the cited work exists; whether the paragraph supports the statement is left to your review.

4.5Limits and work in progress

Further checking rules are planned and will only be described once they are in the program: the version of a provision in the assessment period, format checks of case numbers per court, detection of foreign legal orders and a module for appeals with deadlines computed in code. A benchmark from published decisions is not used as the yardstick – it would come from the same world of sources as the research and would not be independent. Whether an answer is complete and professionally good, the sources alone do not show; for that there is the study mode with the gold standard as an optional scientific add-on – described in the study dossier.

5Re-insertion, answer document and archive

The answer comes back with placeholders. Word or LibreOffice Writer – depending on the setting – shows it with the real data, marked green; “insert real data” or “keep placeholders” is your decision. The answer document is filed unchanged in the archive and used for Word, LibreOffice and PDF; only the HTML view reassembles everything. Every query stays in the archive – grouped by day, with full-text search – as AI assessment, research and log, verification report or all together; the placeholder mapping is stored only locally and readable only by the own user.