{ "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": "2610.20003", "title": "What Coding Agents Leave Unsaid: Measuring Task-State Gaps Between Execution and Reports", "summary": "The method replaces trusting an agent's user-facing report as sufficient evidence of task state with systematic probing that compares what the report conveys against the full execution context. It costs an evaluation harness: recorded execution contexts, a question-generation and answering pipeline over ~37 tasks and thousands of probes, plus annotation or model-based scoring of supervision-relevant states, none of which requires training but does require API inference and engineering time. It can fail if probe questions are unrepresentative of real supervisory needs, if the model-based state assessment misfires on your domain, or if the measured gap does not translate into actionable intervention decisions.", "rationale": "If the reader deploys coding agents in production, this provides a cheap, API-feasible audit of whether agent reports hide failures and workarounds, but the benefit to a specific workflow must be validated before adoption.", "technique": "Task-state recoverability gap measurement for coding agents", "family": "avaliacao_de_agente", "recommendation": "testar", "published": "2026-10-14", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.20003", "independent_implementations": 0, "total_implementations": 1, "stars": 10, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces trusting an agent's user-facing report as sufficient evidence of task state with systematic probing that compares what the report conveys against the full execution context. It costs an evaluation harness: recorded execution contexts, a question-generation and answering pipeline over ~37 tasks and thousands of probes, plus annotation or model-based scoring of supervision-relevant states, none of which requires training but does require API inference and engineering time. It can fail if probe questions are unrepresentative of real supervisory needs, if the model-based state assessment misfires on your domain, or if the measured gap does not translate into actionable intervention decisions.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.20003", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Report-only readers recover 52.0% of task state versus 92.7% for full-context readers, a 40.7-point supervision gap; context-informed report rewrites improve recovery by about 8 points under matched length.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.20003", "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": "" }