{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.20491", "kind": "paper", "url": "https://arxiv.org/abs/2607.20491"}, {"id": "gh:ibm-client-engineering/output-drift-financial-llms", "kind": "repo", "url": "https://github.com/ibm-client-engineering/output-drift-financial-llms"}], "authors": ["Raffi Khatchadourian"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.20491", "source_url": "https://arxiv.org/abs/2607.20491"}, "categories": ["Agents", "Finance", "Language & Knowledge"], "description": {"en": "DFAH-Bench measures whether financial agents reproduce decisions and observable execution paths across qualified replay groups."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.20491", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.20491", "title": "DFAH-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["IBM (United States)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.20491", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.20491", "source_url": "https://openalex.org/W7170919621"}, "key": "claire-radar:2607.20491", "modality": null, "name": "DFAH-Bench", "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_dfah-bench_fb0e51a9", "origin_source": "arxiv", "recognition_confidence": "0.75", "record_sha256": "ba295cf473d69bb9604d010c01e43601dec5bac32093e4343735767b84786960", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T16:18:42.106719Z", "source_url": "https://arxiv.org/abs/2607.20491"}, "publisher": null, "released": "2026-06-10", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.20491", "source_url": "https://arxiv.org/abs/2607.20491"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-20491", "source": "claire_radar", "source_benchmark_id": "2607.20491", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["Finance"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Agent reliability and replayability are active concerns, and the reproducible artifacts support independent inspection.", "score": 6}, "capabilities": [], "confidence": "Medium", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T16:18:42.106719Z", "inputHash": "701e91dd5e3f005a7a5e3a7b83d47204f016f8a77aed905742312bb98a0fd566", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "DFAH-Bench: Benchmarking Observable Agent Instability in Financial Decision-Making", "canonicalNameSource": "paper_title", "decisionReason": "The supplied text defines a named replay protocol, explicit agreement metrics over eligible groups, and a public repository with logs, one-command reproduction, and an interactive explorer.", "evaluationEvidence": {"reuse": "Code, sanitized replay logs, one-command reproduction (make reproduce-paper), and an interactive results explorer: https://github.com/ibm-client-engineering/output-drift-financial-llms", "scoring": "measures decision agreement (DAR) and tool-path agreement (TAR) over the same eligible groups", "task": "The protocol qualifies comparable, sufficiently observed replays and measures decision agreement (DAR) and tool-path agreement (TAR) over the same eligible groups."}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T16:18:42.106719Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "DFAH-Bench measures whether financial agents reproduce decisions and observable execution paths across qualified replay groups.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "evaluation protocol evidence", "public artifact URL"], "snippet": "DFAH-Bench operationalizes the Determinism-Faithfulness Assurance Harness (DFAH), where faithfulness means fidelity of observable execution under replay, not answer correctness."}, "familyId": "bmf_9c5c89f6562b", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_dfah-bench_fb0e51a9", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Financial Services"], "links": {"code": "https://github.com/ibm-client-engineering/output-drift-financial-llms", "data": null, "paper": "https://arxiv.org/abs/2607.20491", "pdf": "https://arxiv.org/pdf/2607.20491", "project": null, "report": "https://arxiv.org/abs/2607.20491"}, "metrics": [], "motivation": "A financial AI agent can repeat a decision while changing the tools, order, or recorded arguments and results used to reach it.", "name": "DFAH-Bench", "oneLine": "DFAH-Bench measures whether financial agents reproduce decisions and observable execution paths across qualified replay groups.", "paperTitle": "DFAH-Bench: Benchmarking Observable Agent Instability in Financial Decision-Making", "primaryDomain": "Finance", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.20491", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "IBM Client Engineering", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/ibm-client-engineering/output-drift-financial-llms"}], "readiness": "Runnable", "recognitionConfidence": 0.75, "relation": "introduces", "releasedAt": "2026-06-10", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Raffi Khatchadourian"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "id": "2607.20491", "title": "DFAH-Bench: Benchmarking Observable Agent Instability in Financial Decision-Making", "type": "arxiv", "url": "https://arxiv.org/abs/2607.20491"}, "sourceUpdatedAt": "2026-07-24T14:53:06Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "It exposes instability hidden by outcome-only evaluation, supporting replay analysis, investigation, and change review."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
