{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.19361", "kind": "paper", "url": "https://arxiv.org/abs/2607.19361"}], "authors": ["Sanjay Mishra", "Divya Chukkapalli", "Ganesh R. Naik"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.19361", "source_url": "https://arxiv.org/abs/2607.19361"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "A session-layer framework for tracking conversational risk accumulation in multi-turn LLM systems, including CRA-Bench datasets and trajectory-native evaluation protocols."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.19361", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.19361", "title": "CRA-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Flinders University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.19361", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.19361", "source_url": "https://openalex.org/W7170143082"}, "key": "claire-radar:2607.19361", "modality": null, "name": "CRA-Bench", "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_cra-bench_d9f4029f", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "09134c183e497a65881f5befd29c91639555776c7893c41f3a7c4bb90bdd67f2", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.19361"}, "publisher": null, "released": "2026-06-03", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.19361", "source_url": "https://arxiv.org/abs/2607.19361"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-19361", "source": "claire_radar", "source_benchmark_id": "2607.19361", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2607.19361", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": [], "confidence": "High", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "5fc97dec5b31864e987507e49e99a3e167e785a4fb449015936fb75a7920c402", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A session-layer framework for tracking conversational risk accumulation in multi-turn LLM systems, including CRA-Bench datasets and trajectory-native evaluation protocols.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "To benchmark CRA, we release CRA-Bench v0.1 (1,200 eight-turn sessions across three threat families with topic-matched benign twins), CRA-Bench v0.2 (LLM-paraphrased variants to reduce template artifacts), and an extended 5-family set (2,000 sessions adding persona priming and context stuffing)."}, "familyId": "bmf_6c4b07513834", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_cra-bench_d9f4029f", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2607.19361", "paper": "https://arxiv.org/abs/2607.19361", "pdf": "https://arxiv.org/pdf/2607.19361", "project": null, "report": "https://arxiv.org/abs/2607.19361"}, "metrics": [], "motivation": "Most safety guardrails for large language models (LLMs) evaluate each prompt-response pair in isolation, which misses failures that arise only over a dialogue as benign turns compose into harm.", "name": "CRA-Bench", "oneLine": "A session-layer framework for tracking conversational risk accumulation in multi-turn LLM systems, including CRA-Bench datasets and trajectory-native evaluation protocols.", "paperTitle": "Stateful Guardrails for Multi-Turn LLM Systems: A Conversational Risk Accumulation Framework", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.19361", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-03", "source": {"authors": ["Sanjay Mishra", "Divya Chukkapalli", "Ganesh R. Naik"], "categories": ["cs.CL", "cs.AI"], "id": "2607.19361", "title": "Stateful Guardrails for Multi-Turn LLM Systems: A Conversational Risk Accumulation Framework", "type": "arxiv", "url": "https://arxiv.org/abs/2607.19361"}, "sourceUpdatedAt": "2026-06-03T02:55:17Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Addresses the gap in evaluating guardrails for multi-turn dialogues where benign turns compose into harm, providing session-level scoring metrics beyond isolated prompt-response checks."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
