{"record": {"aliases": ["Anchor: Mitigating Artifact Drift in Agent Benchmark Generation"], "artifacts": [{"id": "arxiv:2605.26321", "kind": "paper", "url": "https://arxiv.org/abs/2605.26321"}], "authors": ["Abhijay Rana", "Maksim Ivanov"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.26321", "source_url": "https://arxiv.org/abs/2605.26321"}, "categories": ["Agents", "Language & Knowledge", "Manufacturing & Process Control"], "description": {"en": "ERP-Bench evaluates long-horizon procurement and manufacturing tasks in a production-grade ERP system using state-based verification and solver-certified solutions."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.26321", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.26321", "title": "ERP-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2605.26321", "modality": null, "name": "ERP-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_anchor-mitigating-artifact-drift-in-agent-_bde48745", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "5f6bf9968eb1ecfc543155e7388fb09ecfc3083d558de4bc3755d5a297629b5c", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T17:59:55.201635Z", "source_url": "https://arxiv.org/abs/2605.26321"}, "publisher": null, "released": "2026-05-25", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.26321", "source_url": "https://arxiv.org/abs/2605.26321"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-26321", "source": "claire_radar", "source_benchmark_id": "2605.26321", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Anchor: Mitigating Artifact Drift in Agent Benchmark Generation"], "annotation": "Unknown", "applicationDomains": ["Manufacturing & Process Control"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The public generator, enterprise-agent focus, and reported frontier-model performance may attract near-term evaluation interest.", "score": 3}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T17:59:55.201635Z", "inputHash": "7b02f00593fa54db8aead3c51a810b3f60c9f56b2f656a33729797c58564bd99", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We apply Anchor to produce ERP-Bench: a benchmark of 300 long-horizon tasks spanning procurement and manufacturing workflows in a production-grade ERP system.", "canonicalNameSource": "abstract", "decisionReason": "Publish because the abstract formally names ERP-Bench, specifies its task domain, defines a state-based correctness reward, and identifies a public release path for both the generator and dataset.", "evaluationEvidence": {"reuse": "We release the task generator and ERP-Bench dataset at erpbench.ai", "scoring": "rewards depend solely on end-state business correctness", "task": "300 long-horizon tasks spanning procurement and manufacturing workflows in a production-grade ERP system"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T17:59:55.201635Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "ERP-Bench evaluates long-horizon procurement and manufacturing tasks in a production-grade ERP system using state-based verification and solver-certified solutions.", "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 release the task generator and ERP-Bench dataset at erpbench.ai"}, "familyId": "bmf_7bab6c006e10", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_anchor-mitigating-artifact-drift-in-agent-_bde48745", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Manufacturing"], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2605.26321", "pdf": "https://arxiv.org/pdf/2605.26321", "project": null, "report": "https://arxiv.org/abs/2605.26321"}, "metrics": [], "motivation": "AI agents are beginning to complete valuable, long-horizon business operations tasks, but training and evaluation environments for enterprise work still struggle to balance realism, verifiability, and scale.", "name": "ERP-Bench", "oneLine": "ERP-Bench evaluates long-horizon procurement and manufacturing tasks in a production-grade ERP system using state-based verification and solver-certified solutions.", "paperTitle": "Anchor: Mitigating Artifact Drift in Agent Benchmark Generation", "primaryDomain": "Manufacturing & Process Control", "publication": {"evidence": "Accepted to RLEval '26 (Workshop at ACM Conference on AI and Agentic Systems 2026)", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2605.26321", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "RLEval '26 (Workshop at ACM Conference on AI and Agentic Systems 2026)", "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-05-25", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Maksim Ivanov", "Abhijay Rana"], "categories": ["cs.AI"], "id": "2605.26321", "title": "Anchor: Mitigating Artifact Drift in Agent Benchmark Generation", "type": "arxiv", "url": "https://arxiv.org/abs/2605.26321"}, "sourceUpdatedAt": "2026-05-25T20:44:17Z", "topics": ["Agents"], "venueAttempts": [{"decisionRaw": "Accepted to RLEval '26 (Workshop at ACM Conference on AI and Agentic Systems 2026)", "evidence": [{"level": "author-claim", "observedAt": "2026-08-19T11:05:13.395391Z", "rawValue": "Accepted to RLEval '26 (Workshop at ACM Conference on AI and Agentic Systems 2026)", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2605.26321"}], "reviewStatus": "accepted", "venueName": "RLEval '26 (Workshop at ACM Conference on AI and Agentic Systems 2026)"}], "whyItMatters": "It provides auditable enterprise-agent tasks with controlled difficulty, known optimal solutions, and end-state-based rewards."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
