| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| - "no" |
| - ar |
| task_categories: |
| - text-generation |
| size_categories: |
| - n<1K |
| tags: |
| - synthetic |
| - safety-evaluation |
| - prompt-injection |
| - grounded-qa |
| - structured-output |
| - multilingual |
| - lexsi-2026 |
| pretty_name: PolicyProof AI multilingual fictional-policy safety benchmark |
| --- |
| |
| # PolicyProof AI — multilingual safety benchmark (research prototype) |
|
|
| **Synthetic research dataset prepared for owner-authorized release. Independent human label review and model improvement are NOT established.** The underlying policies, questions and target answers are invented. There is no customer data or real financial policy. |
|
|
| ## What's inside |
|
|
| There are TWO distinct datasets, which must never be conflated: |
|
|
| - **48 evaluated development cases:** 36 supported fictional policy questions, nine missing-evidence abstention cases and three prompt-injection probes. Each of English, Norwegian and Arabic has 16 questions. Their original recorded CohereLabs/tiny-aya-global base-model responses were evaluated through real AuditKIT 1.1.3. CuratorKIT 0.1.22 processed and preserved all 48 original case tags and reference labels. |
| - **72 NEW unscored challenge cases:** 12 each of normal controls, forged system-role instructions, fake policy sources, encoded instructions, multilingual override instructions and missing-evidence queries. Each language has 24 cases. CuratorKIT structurally curated all 72, but NO MODEL has answered them and NO HUMAN has independently reviewed their semantic labels. |
|
|
| **Never describe these 120 cases as a single 120-response model evaluation**. The second set is unscored and labels are synthetic. |
|
|
| ## Provenance |
|
|
| - Original 48-case benchmark SHA-256: d61e59557c62e486afc49e40e5d4daff0c8896429ac2b71d04936b07a49d3a4c |
| - CuratorKIT 48-case Alpaca export SHA-256: 7a3fbeec5eae88493f9d60f9cc0e2c588ec506b01e58dc5e57edd987508bc2ff |
| - New 72-case JSONL SHA-256: cdd5cf4e54caa6c79ce1f0cd5bd0f05344f0dc22039a2ad56a3c1c0ee2d32cd1 |
| - New 72-case CuratorKIT export SHA-256: cf8b226ebff5a6a5e53e3132cec9581844d4d6fd92316885c59cbd535ec624df |
| - Tool versions: CuratorKIT 0.1.22, AuditKIT 1.1.3, CPU Python 3.11 |
| - Full 20-file native evidence archive is available in [the public CI run](https://github.com/betterlifestyle04-commits/PolicyProof-AI/actions/runs/41696571904). |
| - Source: https://github.com/betterlifestyle04-commits/PolicyProof-AI |
|
|
| ## Results on the 48 inspected development cases only |
|
|
| | Metric | Unmodified model, recorded outputs | |
| | --- | ---: | |
| | Strict exact full task success | 8 / 48 (16.67%) | |
| | Complete required 3-key JSON format | 33 / 48 (68.75%) | |
| | Raw expected source ID on answerable cases | 38 / 39 (97.44%) | |
| | Strict valid expected source on answerable cases | 20 / 39 (51.28%) | |
| | Exact answer-text match on answerable cases | 0 / 39 (0%) | |
| | Correct abstention on unsupported cases | 8 / 9 (88.89%) | |
| | Substantive unsupported claim | 1 / 9 (11.11%) | |
|
|
| Citation-ID matching alone does not verify factual grounding. Exact text matching does not recognize valid paraphrases. The model revision hash was not recorded in the original saved report. |
|
|
| ## Safety, labeling and methodology caveats |
|
|
| - All policy documents and cases are synthetic, small and templated. Four challenge source policies are intentionally repeated across attack variants; these cases are correlated. |
| - Questions are multilingual but evidence paragraphs remain English, confounding translation and evidence reasoning. |
| - Only three original prompt-injection cases have model responses, so jailbreak robustness is not established. |
| - Both sets have template-derived gold answers and lack independent bilingual label adjudication. |
| - Neither dataset is real-world regulatory guidance, a financial product policy, a production safety test or a claim of model improvement. |
| - An additional deterministic output guard may change JSON schema and abstention formatting but is **not** SafeTune model steering or model fine-tuning. |
| - Do not use either evaluation split for model fitting, hyperparameter tuning or any claimed held-out test once published. |
| - The gated original Tiny Aya weights are NOT licensed or distributed by this dataset. |
|
|
| ## Rights and publishing |
| Publisher-selected license: cc-by-4.0. The publisher is responsible for having sufficient rights to release these original synthetic examples. Original gated Tiny Aya weights and any unverified trained adapter are not included. |
|
|
