{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.29887", "kind": "paper", "url": "https://arxiv.org/abs/2606.29887"}, {"id": "gh:bytedance/safepyramid", "kind": "repo", "url": "https://github.com/bytedance/safepyramid"}], "authors": ["Jiacheng Zhang", "Haoyu He", "Sen Zhang", "Shen Wang", "Xiaolei Xu", "Yuhao Sun", "Meng Shen", "Feng Liu"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.29887", "source_url": "https://arxiv.org/abs/2606.29887"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "SafePyramid evaluates in-context policy guardrailing across 1,000 multi-turn conversations and 3,000 application-specific policies containing 61,699 natural-language rules, organized into three hierarchical capability levels (L0, L1, L2). Scoring is based on violated-rule set prediction with metrics RMR and RDR, and the evaluation harness supports any API or local model."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.29887", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.29887", "title": "SafePyramid"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["The University of Melbourne"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.29887", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.29887", "source_url": "https://openalex.org/W7166743665"}, "key": "claire-radar:2606.29887", "modality": null, "name": "SafePyramid", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_safepyramid_fe239f10", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "865b4145042b04859e3fd4f8cc3318018ad5c477b49402a59c370dc0e55eabd1", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.29887"}, "publisher": null, "released": "2026-06-29", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.29887", "source_url": "https://arxiv.org/abs/2606.29887"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-29887", "source": "claire_radar", "source_benchmark_id": "2606.29887", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/bytedance/safepyramid", "githubScope": "benchmark_repo", "githubStars": 10, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 3, "hfPaperUrl": "https://huggingface.co/papers/2606.29887", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}}, "status": "fresh"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "29270bd355d906f7bc22eec279f2e1c76f1c24b56f454878457e62a4f07f1159", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SafePyramid evaluates in-context policy guardrailing across 1,000 multi-turn conversations and 3,000 application-specific policies containing 61,699 natural-language rules, organized into three hierarchical capability levels (L0, L1, L2). Scoring is based on violated-rule set prediction with metrics RMR and RDR, and the evaluation harness supports any API or local model.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence"], "snippet": "To systematically evaluate this capability, we introduce SafePyramid, a safety benchmark comprising 1,000 multi-turn conversations across 10 domains and 3,000 corresponding application-specific policies, which together contain 61,699 distinct natural-language rules."}, "familyId": "bmf_efe4e93686e2", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/bytedance/safepyramid", "url": "https://github.com/bytedance/safepyramid"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/bytedance/safepyramid", "url": "https://huggingface.co/papers/2606.29887"}], "scope": "benchmark_repo", "url": "https://github.com/bytedance/safepyramid", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_safepyramid_fe239f10", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/bytedance/safepyramid", "data": null, "hfPaper": "https://huggingface.co/papers/2606.29887", "paper": "https://arxiv.org/abs/2606.29887", "pdf": "https://arxiv.org/pdf/2606.29887", "project": null, "report": "https://arxiv.org/abs/2606.29887"}, "metrics": [], "motivation": "In real-world applications, guardrails are often expected to identify unsafe user-model interactions according to application-specific safety policies, rather than relying on predefined risk taxonomies.", "name": "SafePyramid", "oneLine": "SafePyramid evaluates in-context policy guardrailing across 1,000 multi-turn conversations and 3,000 application-specific policies containing 61,699 natural-language rules, organized into three hierarchical capability levels (L0, L1, L2). Scoring is based on violated-rule set prediction with metrics RMR and RDR, and the evaluation harness supports any API or local model.", "paperTitle": "SafePyramid: A Hierarchical Benchmark for In-context Policy Guardrailing", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.29887", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "ByteDance", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/bytedance/safepyramid"}], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-29", "source": {"authors": ["Jiacheng Zhang", "Haoyu He", "Sen Zhang", "Shen Wang", "Xiaolei Xu", "Yuhao Sun", "Meng Shen", "Feng Liu"], "categories": ["cs.AI"], "id": "2606.29887", "title": "SafePyramid: A Hierarchical Benchmark for In-context Policy Guardrailing", "type": "arxiv", "url": "https://arxiv.org/abs/2606.29887"}, "sourceUpdatedAt": "2026-06-29T07:27:22Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "SafePyramid addresses the gap in evaluating guardrails under application-specific policies rather than fixed taxonomies, providing a structured test for rule understanding, dependency resolution, and adaptation to novel frameworks. It enables direct comparison of frontier LLMs and configurable guardrails on a realistic safety task."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
