{"record": {"aliases": ["BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems"], "artifacts": [{"id": "arxiv:2606.20668", "kind": "paper", "url": "https://arxiv.org/abs/2606.20668"}], "authors": ["Leonhard Waibl", "Felix Michalak", "Hadrien Mariaccia"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.20668", "source_url": "https://arxiv.org/abs/2606.20668"}, "categories": ["General AI", "Language & Knowledge", "cs.CR"], "description": {"en": "An operational benchmark comparing moderation filters, jailbreak detectors, and general-purpose LLM supervisors across effectiveness, false positives, latency, and cost."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.20668", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.20668", "title": "BELLS-O"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Graz"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.20668", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.20668", "source_url": "https://openalex.org/W7165652675"}, "key": "claire-radar:2606.20668", "modality": null, "name": "BELLS-O", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_bells-o-evaluating-the-operational-trade-o_538ffeb9", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "3f44cfcb40a1a847b57d24cfa3f0b7afdd512194ee6654a3ae8b58c21ab2d12b", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T18:30:20.889879Z", "source_url": "https://arxiv.org/abs/2606.20668"}, "publisher": null, "released": "2026-06-12", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.20668", "source_url": "https://arxiv.org/abs/2606.20668"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-20668", "source": "claire_radar", "source_benchmark_id": "2606.20668", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems"], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Its coverage of prominent safeguards and deployment costs gives it broad near-term relevance to AI safety practitioners.", "score": 78}, "capabilities": [], "confidence": "Medium", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T18:30:20.889879Z", "inputHash": "2d980891f65abe3f87a94d7e0c15906e85c24d67356ec6dfb30d3435cccef072", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems", "canonicalNameSource": "paper_title", "decisionReason": "The supplied abstract formally names the benchmark, specifies its evaluated systems and comparative metrics, and explicitly releases the benchmark, framework, leaderboard, and datasets.", "evaluationEvidence": {"reuse": "We release the benchmark, framework, leaderboard, and datasets", "scoring": "jointly on detection rate, false-positive rate, latency, and monetary cost", "task": "BELLS-O evaluates 28 systems from 17 providers"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T18:30:20.889879Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "An operational benchmark comparing moderation filters, jailbreak detectors, and general-purpose LLM supervisors across effectiveness, false positives, latency, and cost.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems, Operational), the first independent operational benchmark of LLM supervision systems."}, "familyId": "bmf_6b6cc983d76d", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_bells-o-evaluating-the-operational-trade-o_538ffeb9", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2606.20668", "pdf": "https://arxiv.org/pdf/2606.20668", "project": null, "report": "https://arxiv.org/abs/2606.20668"}, "metrics": [], "motivation": "LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs.", "name": "BELLS-O", "oneLine": "An operational benchmark comparing moderation filters, jailbreak detectors, and general-purpose LLM supervisors across effectiveness, false positives, latency, and cost.", "paperTitle": "BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems", "primaryDomain": "General AI", "publication": {"evidence": "Accepted at the ICML 2026 Workshop on Trustworthy AI for Good (AI4GOOD). 2 figures; main text plus appendices", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2606.20668", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "ICML 2026 Workshop on Trustworthy AI for Good (AI4GOOD)", "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-06-12", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Leonhard Waibl", "Felix Michalak", "Hadrien Mariaccia"], "categories": ["cs.CR", "cs.AI", "cs.LG"], "id": "2606.20668", "title": "BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems", "type": "arxiv", "url": "https://arxiv.org/abs/2606.20668"}, "sourceUpdatedAt": "2026-06-12T07:25:12Z", "topics": ["cs.CR"], "venueAttempts": [{"decisionRaw": "Accepted at the ICML 2026 Workshop on Trustworthy AI for Good (AI4GOOD). 2 figures; main text plus appendices", "evidence": [{"level": "author-claim", "observedAt": "2026-08-19T11:05:13.395391Z", "rawValue": "Accepted at the ICML 2026 Workshop on Trustworthy AI for Good (AI4GOOD). 2 figures; main text plus appendices", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2606.20668"}], "reviewStatus": "accepted", "venueName": "ICML 2026 Workshop on Trustworthy AI for Good (AI4GOOD)"}], "whyItMatters": "It supports deployment-oriented comparison of supervision systems across safety performance and operational constraints."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
