Apertus-Recht: Art. 261bis StGB minimal-twin scenarios (DE / FR / IT)

Content warning. To test where Art. 261bis draws its lines, many items describe discriminatory or degrading statements about real groups in substance (for example, a group called second-class people, a religious group called vermin); no slur is spelled out in the items. The court-validation records quote public decisions, including slurs and insults. Research data, not legal advice.

1980 scenarios (660 per language in DE, FR and IT) about the Swiss offence of discrimination and incitement to hatred, Art. 261bis StGB / CP (Diskriminierung und Aufruf zu Hass); the same scenarios retold as a first-person paraphrase by a user asking about their own conduct, in all three languages (user_voice, language codes de_lay, fr_lay, it_lay; 660 each; the German acts keep the statute's vocabulary, the French and Italian ones (A16) describe the act in everyday words); and the 162-item German exploration battery on which the stage and build columns were measured. Every scenario belongs to a twin group: a base case and variants that each change exactly one legal fact (public ↔ private setting, protected ↔ unprotected characteristic, before ↔ after the 1 July 2020 extension, before ↔ after entry into force in 1995) or one irrelevant detail (placebos: family wording of a public statement, a synonym of the characteristic, a date that does not matter). The gold verdict is computed by code from the pinned statute text and BGE 130 IV 111 E. 5.2, never written by hand.

Dataset Details

• Curated by: the Apertus-Recht team (Hack Apertus 2026, Track 1B, Swiss Voices › Behavioural Alignment).
• Language(s): German, French, Italian (Swiss legal register; the official statute terms of each language).
• License: CDLA-Permissive-2.0 (all items; per-item licence components and provenance in metadata/items.jsonl for the core and first-person sets, and in metadata/items.jsonl inside dignity/, paraphrase/ and stgb_179ter/).
• Git repository: https://github.com/mi11ione/apertus-recht (generator, gold engine, harness, tests; Apache-2.0).
• Version: 1.8.5 (generator 1.0.0, seed 261). Changelog at the end.

Uses

Direct Use

Evaluating whether a language model applies the Swiss law in force when a user asks a local legal-information assistant whether a statement is punishable under Art. 261bis. The model answers each scenario at seven prompt rungs (R0: plain question; R1: elements named; R2: + a one-sentence paraphrase of BGE 130 IV 111 and the list; R3: + "the list is closed" and that gender identity and sex are not covered; R4: + "sexual orientation only from 1 July 2020"; R5: the verbatim Fedlex text of Art. 261bis and Art. 2 StGB with Fedlex's footnote and the Regeste of BGE 130 IV 111; R6: R5 with an instruction to check each element) in two answer framings (Ja/Nein; "Urteil: schuldig / nicht schuldig", and the French and Italian equivalents). The twins read back which element a model gets wrong, not only how often. The same items support the two other behavioural-alignment measurements: preference accuracy (forced choice between each item's chosen and rejected answer) and free generation judged against the item's acceptance criterion (scripts/judge_open.py in the repository: two blind Claude judges, a judge of another model family (Qwen), a deterministic parser, every judgement hash-bound to the answer it judged, and a seeded 20% sample laid out so that anyone can audit the judges).

The dignity threshold (A18)

The dignity threshold is the coded ground of 6 of the 23 coded court acquittals, the most frequent single ground (three of them statement-based Abs. 4 acquittals, one not final): 20 twin groups per language in which a public statement about a religious group is either degrading (denying, or at least calling into question, the believers' equal worth as human beings: convict), a criticism of the religion's institutions or teachings that stays objective and cites an objective basis (a published report, the institutions' own statements or texts; acquit: BGE 131 IV 23 E. 3, Regeste pinned in DE/FR/IT), or a crude joke whose wording the facts do not give (open, a modelling choice: a court would acquit if the wording is not established, Art. 10 Abs. 3 StPO, and the criterion accepts that answer). One of the four criticisms (no religious instruction in public schools, citing the state's confessional neutrality) is a policy demand (act a4D_crit4 in slots.act); it has its own answer and criterion, resting on the restraint BGE 131 IV 23 E. 3.1, 3.3 and 3.4.2 asks for in political debate. The "opaque finances" criticism is never paired with Judaism (A22). Gold comes from the element dignity in the statute model; reviewed by a legal-reviewer agent before release. Files dignity/eval/ (items, prompts with the BGE 131 IV 23 Regeste at the verbatim-law rungs) and dignity/responses/; results in the repository (reports/dignity.json). Scope: the degrading variant uses the Regeste's own formula, so it marks the clear ends of the line; Art. 261 StGB (disturbing freedom of belief) is out of scope. Exploratory, not pre-registered. Read against 7 court rulings (A20, review/dignity_court.json; 3 acquittals on the dignity threshold, 4 convictions): each case summarised as facts only, the element value coded blind from the Regeste of BGE 131 IV 23 by two Claude agents and Qwen-SEA-LION. Every coder codes all 4 convictions degrading. Agreement with the courts: 6, 6, 4 of 7 (always-convict: 4); the two Claude coders agree with each other on 7 of 7. (These are the codings after A22, which corrected the coder instructions and completed one summary; the round before, on eight rulings, gave 5, 5 and 4 of 8. A24 removed a first-instance acquittal still under appeal.) The one acquittal both Claude coders code degrading was a first-instance conviction reversed on appeal, partly on the subjective side. Each coding concerns a facts-only summary, not the persons; the acquittals stand, and the coders' written reasons are not released. No sampled ruling concerns criticism of a religion's doctrine or institutions, so the check is consistent with the gold here but does not test it.

Natural paraphrases (A16)

To test whether the findings depend on the templated wording, a stable sample of 40 twin groups (the same in DE, FR and IT; 120 items per language) was rewritten in everyday genres (news brief, witness account, police journal entry, forum post, social worker's note, letter to the editor) with the legal facts locked. Every paraphrase was read blind by a Claude agent and by Qwen-SEA-LION-v4-32B; it is kept only if the gold computed from each reader's extraction equals the item's gold (verdict and deciding element) and the date, act, severity and intent match: 334 of 360 kept (paraphrase/eval/paraphrase_<lang>.jsonl, each with paraphrase_of, genre and the original's gold; prompts and the 70B's responses alongside; verification in the repository, reports/paraphrase_verification.json). The 70B behaves on them as on the templated originals (reports/paraphrase.json). Exploratory, not pre-registered.

Out-of-Scope Use

Not legal advice and not a model of every element of the offence: intent is fixed, the Abs. 4 dignity threshold is written to be met outside the dignity family (A18), Abs. 3 and 5 are not covered, so the items measure how a model applies the elements, not whether it predicts a conviction. The free-text question asks whether the act is punishable under Swiss law, while the gold is Art. 261bis; every answer judged misleading names Art. 261bis as its basis. The Italian verdict framing uses "proscioglimento", which is wider than an acquittal on the merits ("assoluzione"). Not a training set: the canary string in every item record (the eval/ item files; the prompt files carry the scenario text without it, so filter them by item id) marks it as evaluation data, to be kept out of training corpora. Do not train on generator output either: the reference answers are written per legal reason, so most of them recur verbatim for any seed. chosen_with_scope documents the answer an assistant should give; use it as a specification for training data written independently. Content warning: see the top of this card.

Specifications

• Use case → specification: the question a person asks about their own or someone else's speech; the elements the answer turns on (publicity, characteristic, date) are exactly the slots the twins vary.
• Languages follow where the article is litigated: of the decisions that cite BGE 130 IV 111, most are French- and German-language (court-validation pool: 67 DE, 79 FR, 4 IT decisions).
• Localisation: each language is set where it is spoken: German items in German-speaking towns, French items in Romandie (Lausanne, Genève, Sion, …), Italian items in Ticino and Italian Graubünden (Lugano, Bellinzona, Poschiavo, …), with each language's official legal terms and citation forms (StGB / CP, BGE / ATF / DTF, AS / RO / RU).
• Prompt rungs model the retrieval levels of an assistant: R0 no law; R1–R4 the elements, a one-sentence paraphrase of BGE 130 IV 111, the closed list and the 2020 date as rules; R5 the verbatim Fedlex text of Art. 261bis (with Fedlex's footnote) and Art. 2 StGB plus the Regeste of BGE 130 IV 111, as a retrieval pipeline would paste them.
• Acceptance criteria, chosen / rejected answers, settled / contested labels and a normative source are given for every item (Track 1B behavioural-alignment requirements).

Dataset Structure

eval/core_<lang>.jsonl (one row per scenario), eval/battery162.jsonl, eval/prompts_core_*.jsonl (one row per scenario × rung × framing: the exact user message, the assistant prefix and the two candidate answers), eval/prompts_open_core_*.jsonl (free-text question at R0, R3 and R5, for generation and judging against the acceptance criterion), eval/prompts_pref_core_*.jsonl (settled items: the item's chosen and rejected answers as the two candidates, stored by direction, with chosen_side; preference accuracy), metadata/items.jsonl, metadata/sources.json, metadata/generator/ (the YAML the items were generated from), responses/ (model responses, one file per model build), MANIFEST.json (sha256 of every file).

Main fields: id, lang, twin_group, family, variant (base / true_flip / placebo / contested), changed_slot, base_id, slots (template ids), facts (the legal facts), text, gold_verdict (convict / acquit / open), deciding_element, gold_reason, label (settled / contested), trap, acceptance_criterion, chosen (the answer the law supports, specific to the item), rejected (a wrong answer phrased as models argue, for preference data), rejected_minimal (settled items: the chosen answer mirrored, same wording and length, verdict and deciding claim negated; the preference prompts use this one, because a shorter rejected lets a likelihood comparison measure length), scope_note and chosen_with_scope (every Art. 261bis item of the core, first-person, dignity and paraphrase sets; not the 162-item exploration battery or the Art. 179ter set; A17, A19, A21: on acquittals what "not punishable under Art. 261bis" does not mean and where else to turn; on convictions and open items that this is an assessment of the stated facts, not legal advice, and that a court also weighs intent and evidence; chosen_with_scope is the answer an assistant should give, chosen stays the scored reference), rationale (the element-by-element derivation), citation, normative_source, split (fit / heldout wording family), licence, canary.

family (DE count)	what changes	twin groups
SET	public → private setting; family-worded public placebo	55
CHAR	protected → unprotected characteristic; synonym placebo	60
CHARX	protected → two unprotected characteristics	20
DATE_SO	sexual orientation 2021–25 → before 1 Jul 2020; boundary placebo (1 Jul / Aug 2020)	35
DATE_CTRL	race/ethnicity/religion 2021–25 → 2018–20 (placebo: still protected) and → 1993/94 (before entry into force)	25
CONTESTED	public → fact-dependent setting (open); family-worded placebo	15
WORDING	other wording of the same paragraph; another paragraph	10

Per language: 220 base, 215 true flips, 210 placebos, 15 contested. Verdicts: 430 convict, 215 acquit, 15 open. Characteristics: sexual_orientation 175, ethnicity 154, religion 137, race 94, sex 18, age 17, gender_identity 16, canton 16, occupation 16, disability 9, political_opinion 8. Paragraphs: abs1 220, abs2 218, abs4 222. Diversity (DE): 8 frames, 18 act wordings in 3 wording families, 29 settings, 33 group phrases, 14 dates, 12 places, 8 actors; mean 44.4 words; distinct-2 0.0516, distinct-3 0.0981 (templated by design).

Second statute: Art. 179ter StGB (stgb_179ter/, exploratory)

The same generator, gold engine, QC gates and harness on a different kind of offence: recording a non-public conversation one takes part in, without the other participants' consent (Art. 179ter StGB), prosecuted only on a complaint filed within three months of learning of the act and the offender (Art. 30 and 31 StGB). Five elements: non-public conversation, the recorder is a participant (a third party falls under Art. 179bis instead), sound recording, no consent, timely complaint (a procedural element 261bis does not have). 120 German scenarios in 40 twin groups (COMPLAINT 32, CONSENT 24, CONV 24, MEDIUM 24, ROLE 16); verdicts 72 convict, 48 acquit. Sources pinned like 261bis: the Fedlex texts of Art. 179ter, 30 and 31 StGB (version of 1 Oct 2026, sha256 in stgb_179ter/metadata/sources.json), and, for the scope notes the answers make, Art. 179bis and 179quinquies StGB and Art. 329 StPO (A15). Acts are dated from September 2023, after the harmonised penalty ranges entered into force (1 Jul 2023). Prompts at three rungs (R0 no law; R1 the elements named; R2 plus the complaint rule) in the two framings; every question asks about a conviction under Art. 179ter alone, since a court is not bound by the legal qualification of the charge. German only, not pre-registered and not part of the held-out test: it shows that the method carries to a second statute, not a second headline. A legal-reviewer persona (a Claude agent) found all 120 gold labels correct and asked for wording fixes, applied before any model was scored: the question restricted to the article, a police report instead of an indictment (no indictment of a complaint offence without a complaint), no legal conclusions in the facts ("Täter", "betroffen"), a rent talk in the landlord's flat instead of a phone call (business phone calls on orders, contracts, reservations and similar are exempt under Art. 179quinquies Abs. 1 lit. b). Scope of the gold: it answers only "conviction under Art. 179ter". An outsider who records may be punishable under Art. 179bis; without a timely complaint the proceedings are not opened or are discontinued (Art. 310 Abs. 1 lit. a, Art. 319 Abs. 1 lit. d, Art. 329 Abs. 4 StPO), so for the 16 complaint items the verdict framing has no correct answer and they are read in the Ja/Nein framing only; this is not an acquittal on the merits; justification and Art. 179quinquies are assumed not to apply, intent is assumed present; a video with sound counts as a sound recording (only handwritten notes are tested as a medium that is not); «schuldig / nicht schuldig» are answer tokens, not Swiss dispositive wording. Files: stgb_179ter/eval/, stgb_179ter/metadata/, stgb_179ter/responses/, stgb_179ter/MANIFEST.json; results in the repository (results/results_179ter.json).

Apertus v1.5 70B (publicai-70b-179ter):

rung	framing	acquittals recognised	convictions kept	balanced accuracy
R0	Ja/Nein	48/48	0/72	0.50
R0	schuldig / nicht schuldig	3/48	72/72	0.53
R1	Ja/Nein	11/48	72/72	0.61
R1	schuldig / nicht schuldig	0/48	72/72	0.50
R2	Ja/Nein	34/48	58/72	0.76
R2	schuldig / nicht schuldig	22/48	72/72	0.73

acquittal decided by (Ja/Nein)	R0	R1	R2
complaint:failed	16/16	0/16	16/16
consent:failed	8/8	3/8	7/8
conversation:failed	8/8	0/8	2/8
medium:failed	8/8	8/8	6/8
role:failed	8/8	0/8	3/8

A framing that answers one way for every scenario reads no element: with no law (R0), Ja/Nein says "Nein" to all 120 (all acquittals "recognised", no conviction kept), while the verdict word convicts all. Only with the complaint rule stated (R2) do both framings separate the cases; the public setting and the outsider are still convicted there.

Dataset Creation

Curation Rationale

Apertus is trained on Swiss court decisions, 147 of which cite BGE 130 IV 111, yet plain-chat answers convict for statements made only within the family and for characteristics outside the closed list (see the technical report). A minimal-twin set with computed gold separates which element of the law a model applies from how often it is right, and makes each wrong element measurable across training stages and model builds.

How it differs from existing datasets (sources and details: docs/prior-art.md in the repository). Swiss legal benchmarks (LEXam, SwissLegalEvals, Swiss Judgment Prediction, SCALE) score exam answers or judgment outcomes, with no one-fact variants and no statute versions. LawShift (Chinese) pairs minimal case edits with hypothetical statute revisions graded by experts; TALD (Chinese) asks which statute version applies, without twins; OpenExempt (US) computes gold with a solver but has no amendment dates; Multilingual HateCheck has templated contrasts but no law. To our knowledge this is the first set that combines one-legal-fact twins and placebos, real amendment dates with non-retroactivity (1 Jan 1995, 1 Jul 2020, Art. 2 StGB), gold computed from pinned official texts and checked against court decisions, in German, French and Italian.

Source Data

Normative sources only (all public domain, Swiss Copyright Act Art. 5): Art. 261bis StGB in the versions before and since 1 July 2020 and Art. 2 StGB (Fedlex ELI eli/cc/54/757_781_799, DE/FR/IT, sha256-pinned, metadata/sources.json), the amendment footnotes (AS 1994 2887; AS 2020 1609), the Regestes of BGE 130 IV 111 and BGE 131 IV 23 (DE/FR/IT), and for the second statute Art. 179bis, 179ter and 179quinquies StGB and Art. 329 StPO. Court decisions for validation (public domain, URG Art. 5): voilaj/swiss-caselaw (CC0) and, for the dignity check (A20), entscheidsuche.ch; referenced by id and URL only.

Data Collection

No data was collected from people. Scenarios are generated from YAML templates (metadata/generator/) by a seeded, deterministic generator: statute-twins build reproduces every generated file byte for byte (checked by make run). The natural paraphrases (A16) are the exception: written by Claude agents from each item's facts, verified blind, and stored with their verification (provenance.text_source).

Source Data Producers and Data Subjects

Templates: written by the team with Claude Code agents (German), translated by Claude agents (French, Italian), reviewed as described below. Every "reviewer", "juror", "counsellor" or "lawyer" named in this card and in the amendments is a Claude agent given that role, not a person. The scenarios have no data subjects: actors are generic ("a man", "a student"), places are cities, groups are described by their characteristic. The court-validation records are the exception: they hold pseudonymised data on criminal proceedings from public, anonymised decisions (docket, URL, outcome, coded elements, notes that name no private person), released for verification.

Ownership and Consent Management

The team owns the generated text and releases it under CDLA-Permissive-2.0. The scenarios contain no personal data, and no human annotators or reviewers took part (all labelling steps are reproducible agent runs). Anyone who extends the set with human reviewers or human-written scenarios uses the consent and AI-use form in docs/consent-template.md (pseudonymised reviewers, signed forms kept by the maintainers, withdrawal before each release). AI permissions: generated with AI assistance (declared). The licence permits any use. The set is built for evaluation: we ask that it is kept out of training data (the canary string in every item record lets corpus builders exclude it), and that chosen_with_scope serves as a specification for independently written training data, not as data.

Data Processing and Quality Control

Automated gates (statute-twins qc, all passing for this release): schema; gold recomputed from the facts for every item; twin invariants (each variant differs from its base in exactly one slot; placebos keep and true flips change the verdict); duplicates (0 in the core set); language identity; PII scan (Swiss phone, AHV, IBAN, e-mail, addresses, name gazetteer) over items, responses, metadata, the released review records and the dignity and paraphrase splits: 0 hits; canary. The generator refuses to build if a variant changes more than one slot, if the gold engine's role differs from the design, or if a month-precision date's verdict depends on the day. Statute tests check the closed-list terms and the amendment dates against the official Fedlex text and footnotes, and mutation tests check that changing one rule flips exactly the items that rule governs.

• Translation: FR/IT by Claude agents with the official CP terms; blind back-translation into German by separate agents; comparison for legal facts: 0 critical, 3 minor findings, all fixed. A localisation review by a French- and Italian-speaking lawyer persona then fixed what back-translation cannot see (German places, calqued intent formula and verdict words, citation form); amendment A7, before any French or Italian result was analysed.
• Contamination (scan of 9 October 2026 on the core items as they stood then, before later wording amendments; the first-person, dignity and paraphrase splits were written later and not scanned) against voilaj/swiss-caselaw (listed as v1.5 pre-training data in the Apertus data catalogue, github.com/swiss-ai/apertus-data-transparency; revision c4158e6; 1,069,365 decisions, all 121 data files including GE, VD and BVGer): of 35,241 distinct word 13-grams of our items, 0 scenario 13-grams occur in the corpus (0 statute-derived). Exposure of the law itself: 147 decisions mention BGE 130 IV 111, 754 mention Art. 261bis.

Data Instance Licenses

Every item: CDLA-Permissive-2.0 (original text). Components cited by each item (statute, amendment, BGE) are public domain under URG Art. 5; listed per item in metadata/items.jsonl (licence_components). Model responses in responses/ are outputs of Apertus (Apache-2.0 model licence) and Qwen3 (Apache-2.0); the judgements and annotations in review/ are outputs of Claude agents (Anthropic; outputs owned by the team under Anthropic's terms) and of Qwen-SEA-LION-v4-32B (a Qwen3-based model served through Public AI); all are released under CDLA-Permissive-2.0.

Annotations

Annotation Process

Gold labels are computed, not annotated. Annotation checks them: 960 items (all 660 German items and a stratified sample in FR and IT) were labelled blind (opaque ids, no access to the code or the gold) by six Claude agents following docs/annotation-guidelines.md. Agreement with the code gold on settled items: 939/939; deciding element on acquittals: 315/315; Cohen's κ (3 classes, all items) 1.000. In the first round 7 of 21 contested items were labelled with a verdict and one contested setting was read as clearly private by 5/5 annotators; that setting and an anachronism the annotators found (acts in 1993/94 posted online) were fixed before any model result was analysed (pre-registration amendment A3) and the changed items re-annotated. A second blind pass with new opaque ids gave the annotators only the law (statute texts, Art. 2 StGB, the Regeste of BGE 130 IV 111; no rule summary): 939/939 settled items agree with the code gold, κ against gold 1.000. The same law-only task given to a model of another family (Qwen-SEA-LION-v4-32B, amendment A15): 910/939 settled items agree (κ 0.93); each of its 29 disagreements convicts a case the pinned text does not cover (sexual orientation before 1 July 2020, acts before 1995, an unlisted characteristic), the errors Apertus makes, not gold errors. All labels: review/. Adjudication of every annotator disagreement, every contested item and a blind 10% sample of settled items by two independent agent adjudicators, 115 items (blind sample 94, contested 21, annotator disagreements 0); adjudicators agree on 115/115; gold wrong: 0/94 in the blind sample of settled items (Wilson 95% upper bound 3.9%), 0/21 contested items. Court validation of the gold engine against real decisions: 326 decisions triaged (stratum A: 150, stratum B: 176); 57 in scope (55 convictions, 2 acquittals); engine agrees with the court in 57/57 (Wilson 95% 0.94-1.00; an engine that always convicts would agree in 55/57; acquittals 2/2, Wilson 0.34-1.00); 24 out of scope (unmodelled grounds), 50 unextractable, 195 excluded; every tuple extracted twice by independent agents, conflicts settled by a third (amendment A4). Counted per distinct case (post hoc, amendment A12: repeat instances of one case merged; the Perinçek conviction, 6B_398/2007, left out because the ECtHR found a violation of Art. 10 ECHR and the Federal Court revised it, 6F_6/2016): 52/52, 6 of them genocide denial; of the 2 acquittals one is clean and one agrees on the outcome on another ground (amendment A15 reclassified BGE 126 IV 176, a cassation and remand, as not an acquittal).

Annotator Details

Annotation, adjudication and court-tuple extraction were done by Claude agents (Anthropic), declared as AI; a model of another family (Qwen-SEA-LION-v4-32B) annotated, judged and coded independently. Annotators worked blind (opaque ids, no access to code or gold); adjudicators and legal reviewers then checked the gold itself. No human labels: every label in this release can be regenerated by anyone with the same instructions (docs/annotation-guidelines.md and the scripts in the repository). The agents share a model family with the generator's author, so their agreement rules out reading errors, not a shared misconception of the law; the statute tests against the official text and the court validation check the law itself.

Risks, Bias, and Limitations

• One statute; templated scenarios (low lexical diversity by design); intent is fixed, and so is the dignity threshold outside the dignity family (A18).
• Contested items are few (15 per language) and their verdict is open by construction.
• FR/IT are translations, checked by blind back-translation and a localisation review, not written by native lawyers.
• Model responses come from specific builds; the Public AI endpoint named swiss-ai/apertus-v1.5-8b serves the 70B (evidence in the repository's docs/builds.md); its rows are labelled 70B.

Personally Identifiable Information

Items: none (no names, no identifiers; PII gate result above). Validation records cite public court decisions by docket and URL; their notes name no private person (only case names such as Perinçek v. Switzerland), and the fact summaries written for the dignity check are withheld (pseudonymised data on criminal proceedings): only their sha256, the docket, the court outcome and the coders' values are released.

Sensitive Information

The scenarios mention protected characteristics (race, ethnicity, religion, sexual orientation, gender identity, etc.) of groups, as the statute requires; no individual's characteristic is described. Some items state degrading content in substance (see the content warning at the top of this card); no slur is spelled out.

Recommendations

Score with both framings and read the element table, not a single accuracy; report the build (precision, conversion) alongside results, since 4-bit conversions changed verdicts in our measurements.

Model responses in this release

37 runs (responses/<run>.jsonl with <run>.meta.json): 8b-bf16-DPO, 8b-bf16-RLVR, 8b-bf16-blend-RLVR-DPO-a0.25, 8b-bf16-blend-RLVR-DPO-a0.5, 8b-bf16-blend-RLVR-DPO-a0.75, 8b-gguf-q4_k_m, 8b-gguf-q4_k_m-cpu-core, 8b-gguf-q4_k_m-cpu-core-fr, 8b-gguf-q4_k_m-cpu-core-it, 8b-gguf-q4_k_m-cpu-pref, 8b-gguf-q8_0, 8b-mlx4-DPO, 8b-mlx4-RLVR, 8b-mlx4-SFT, 8b-mlx4-base-completion, 8b-mlx4-final, publicai-70b-ctrl-de, publicai-70b-ctrl-fr, publicai-70b-ctrl-it, publicai-70b-fc, publicai-70b-fc-core, publicai-70b-fc-core-fr, publicai-70b-fc-core-it, publicai-70b-fc-lay, publicai-70b-fc-lay-fr, publicai-70b-fc-lay-it, publicai-70b-gen, publicai-70b-open, publicai-70b-open-fr, publicai-70b-open-it, publicai-70b-pref, publicai-70b-pref-fr, publicai-70b-pref-it, publicai-70b-r5x-de, publicai-70b-r6-fr, publicai-70b-r6-it, qwen3-8b-mlx4.

Changelog

• 1.8.5 (2026-10-11): amendment A27: wording corrections; three reference answers refined (policy demand, borderline joke, first-person re-posting). No prompt or result changed.
• 1.8.4 (2026-10-11): amendment A26: report restructured; court-check summaries withheld (hashes and codings stay); precision in the card and answers. No prompt or result changed.
• 1.8.3 (2026-10-11): amendment A25: figure encoding (undetermined cells), legal precision in answers and notes, card wording. No prompt or result changed.
• 1.8.2 (2026-10-11): amendment A24: the dignity court check reads seven final rulings (a ruling under appeal removed); R4 and dignity paired statistics; wording of coverage and claims. No prompt or result changed.
• 1.8.1 (2026-10-11): amendment A23: BGE 131 IV 23 considerations pinned; policy-demand answer and note; one scope note per open reason; wording of court baselines and the dignity court check. No prompt or result changed. The names removed from the court notes in 1.8.0 were removed from the whole git history too, and the Hugging Face history was squashed.
• 1.8.0 (2026-10-11): amendment A22: dignity criticisms with an objective basis (re-scored), the policy demand with its own answer; victim routes on punishable items; five names removed from court notes and the PII gate extended to all review files; repaired controls and statistics (review/controls_a22.json, review/stats_a22.json). The git tags v1.6.0 to v1.8.3 no longer pass the dignity court step (selfcheck 2i): the court-check files they needed were removed from the history (A23, A26); v1.8.4 and later reproduce in full.
• 1.7.0 (2026-10-11): amendment A21: scope notes on every item; BGE 131 IV 23 stated in full in the gold answers; first-person time-bar note corrected; per-item licence and provenance for the dignity and paraphrase splits; content warning; consent template completed; A20 re-coded on redacted case summaries. No prompt or response changed.
• 1.6.0 (2026-10-11): amendment A20, the dignity element checked against eight court rulings coded blind (review/dignity_court.json); no item, prompt or result changed.
• 1.5.0 (2026-10-11): amendment A19, case-specific scope notes (referral by characteristic, limits of honour and personality claims, time bars, criticism as protected speech, speaker-facing notes for first-person items); the dignity answers restate BGE 131 IV 23's test; no prompt or result changed.
• 1.4.0 (2026-10-11): amendment A18, the dignity threshold as an element and the DIGNITY family (60 items per language, dignity/); no core item, prompt or result changed.
• 1.3.0 (2026-10-11): amendment A17, a scope note and chosen_with_scope on every acquittal; no gold, prompt or result changed.
• 1.2.0 (2026-10-11): amendment A16 (first-person retellings in French and Italian, natural paraphrases verified blind by two model families, the 8B in French and Italian); no item or gold label of 1.1.0 changed.
• 1.1.0 (2026-10-10): amendments A12–A15: post-hoc court recount per distinct case, judgements bound to answers and provenance repairs (A12); controls for agreement bias and for R5 (A13, review/controls_a13.json); free text in French and Italian and Art. 179ter as a second statute (A14); corrections, a length control, an other-family gold check (A15); no item or gold label of 1.0.0 changed. Files and hashes: MANIFEST.json (204 files).
• 1.0.0 (2026-10-10): first release. Generator 1.0.0, seed 261. Pre-registration (docs/PREREG.md) was locked before any model scored the core set; every later change is an amendment in docs/AMENDMENTS.md: A1 design moved into YAML, output byte-identical; A2 FR/IT surface fixes from back-translation; A3 one contested setting and an anachronism (online acts in 1993/94) fixed after blind annotation, before any result was analysed; A4 agent verification (double extraction, two adjudicators) instead of a human reviewer; A5 four wording fixes found in review after the German results had been analysed, all affected prompts re-scored, the pre-A5 responses kept in the Git repository (archive/) so the effect of the fix is visible; A6 item-specific reference answers (setting, characteristic, paragraph and date of each scenario; 333 distinct chosen per language), the free-text and preference prompts, and unparsed answers counted in every recall denominator; A7 French and Italian localisation (places in Romandie and Ticino, intent formula, verdict nouns, citation forms), before any French or Italian result was analysed, with both languages re-scored in full; A8 R5 (the verbatim law), length-matched preference pairs (rejected_minimal), Art. 1 and 2 StGB in the answers, a contested criterion that accepts a reasoned leaning, and the court-validation baseline; A9 the German scenarios as a first-person paraphrase (de_lay); A10 R6, the control for R5; A11 German localisation (bus, public information evening), free text with the verbatim law, a Qwen judge.

Citation

@misc{apertus_recht_2026,
  title  = {Apertus-Recht: Art. 261bis StGB minimal-twin scenarios (DE/FR/IT)},
  author = {Team 444},
  year   = {2026},
  note   = {Hack Apertus 2026, Track 1B. Dataset version 1.8.5},
  url    = {https://github.com/mi11ione/apertus-recht}
}

Glossary

• Twin group: a base scenario and its variants; each variant changes exactly one slot of the base.
• True flip / placebo / contested: the statute decides the variant differently from the base / the same / not without facts the scenario leaves open (gold open). Assigned by the gold engine, not by hand.
• Deciding element, gold reason: the element of Art. 261bis that decides an acquittal (setting:failed, characteristic:not_listed, characteristic:not_yet_covered, act_date:not_in_force).
• Rung R0–R6: how much law the prompt contains: R0 none; R1 the elements named; R2 + a one-sentence paraphrase of BGE 130 IV 111 and the list; R3 + "the list is closed"; R4 + "sexual orientation since 1 July 2020"; R5 the verbatim Fedlex text of Art. 261bis and Art. 2 StGB with Fedlex's footnote and the Regeste of BGE 130 IV 111; R6 = R5 with R1's instruction to check the elements (control).
• Framing: conv (Ja/Nein), ver ("Urteil:" schuldig / nicht schuldig; FR condamnation / acquittement, IT condanna / proscioglimento), open (free text), pref (chosen vs rejected_minimal).
• Read-back status: F the model convicts where the statute acquits (element in force in the model); a applied; u unreadable (the unchanged reference is not convicted either); ? undecided. Red: F in every readable framing.
• Wording family A/B/C: three sets of act wordings; A and B fit the read-back, C is held out.

Dataset Card Authors

Team 444 (Hack Apertus 2026), with Claude Code agents as declared above.

Dataset Card Contact

Via the Git repository's issue tracker: https://github.com/mi11ione/apertus-recht.