{ "schema_version": 1, "builder_plan": { "scenario": "An agent looks good in demos, but you cannot tell which product tasks it fails.", "change": "Use the study's evaluation idea to inspect a small set of tasks with human-checked answers.", "baseline": "Your current evaluator and a manually reviewed reference set.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Agreement with human review, missed failures, false alarms, and evaluation cost per task.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2610.31325", "title": "Consistent but Biased: Chair Personas Trade Noise for Bias in Multi-Agent Clinical Decisions", "summary": "The paper does not propose a new method to adopt; it replaces the implicit practices of leaving the chair agent's persona unspecified and of treating raw inter-run agreement as reliability, recommending instead that the chair persona be reported and systematically varied and that agreement be chance-corrected (e.g., Fleiss kappa). The cost is additional evaluation runs across persona conditions and prompt formats (rules versus personality traits), which is inexpensive with API-based models but multiplies the number of simulated deliberations per case. What can fail: persona prompts can silently shift decisions (operate-now rates moved from 2% to 93% across personas while the underlying plan barely changed), personality-trait prompts reproduce decision timing but not safety-relevant behaviors, and uncorrected agreement can be mistaken for reliability when it is no better than chance.", "rationale": "Any team deploying multi-agent LLM pipelines with a deciding agent should treat the decider's prompt persona as a reported experimental variable and chance-correct run-to-run agreement, since unreported personas demonstrably bias outcomes and raw agreement overstates reliability.", "technique": "Chair persona control in multi-agent LLM decisions", "family": "avaliacao_de_agente", "recommendation": "adotar", "published": "2026-10-22", "as_of": "2026-10-23", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.31325", "independent_implementations": 0, "total_implementations": 1, "stars": 10, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "The paper does not propose a new method to adopt; it replaces the implicit practices of leaving the chair agent's persona unspecified and of treating raw inter-run agreement as reliability, recommending instead that the chair persona be reported and systematically varied and that agreement be chance-corrected (e.g., Fleiss kappa). The cost is additional evaluation runs across persona conditions and prompt formats (rules versus personality traits), which is inexpensive with API-based models but multiplies the number of simulated deliberations per case. What can fail: persona prompts can silently shift decisions (operate-now rates moved from 2% to 93% across personas while the underlying plan barely changed), personality-trait prompts reproduce decision timing but not safety-relevant behaviors, and uncorrected agreement can be mistaken for reliability when it is no better than chance.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.31325", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }