{ "schema_version": 1, "builder_plan": { "scenario": "Your coding assistant makes changes that still need substantial repair before merging.", "change": "Try the technique on held-out issues from a repository with reliable tests.", "baseline": "The same coding agent, repository snapshot, tools, and token budget.", "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": "Correct patches, regression failures, human repair time, and total cost per accepted change.", "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": "2610.18845", "title": "The Last Remote Sensing Model Built by Humans? Casting Any Earth Observation Task as Recursive Self-Improvement", "summary": "It replaces hand-designed EO training recipes and per-task expert tuning with a general-purpose coding agent that, given only a harness and a verifier, builds the model and recursively improves it against the verifier. The cost is sustained agent API spend plus many local training and evaluation runs on the single GPU, together with the engineering effort to construct a verifier and harness that genuinely reflect the production objective. It can fail by overfitting or gaming the verifier, leaking benchmark information through its own search, emitting non-reproducible artifacts, or running unbounded improvement loops that exhaust API and compute budget without off-benchmark gains.", "rationale": "The approach runs through third-party agent APIs and trains task models small enough for a 24 GB GPU, so it is within reach, but its state-of-the-art and label-efficiency claims are benchmark-specific and the recursive loop carries verifier-gaming and budget risks that require validation on the reader's own EO workload before production adoption.", "technique": "Harness-verifier recursive coding agent for EO", "family": "agentes_de_codigo", "recommendation": "testar", "published": "2026-10-13", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.18845", "independent_implementations": 0, "total_implementations": 1, "stars": 9, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "It replaces hand-designed EO training recipes and per-task expert tuning with a general-purpose coding agent that, given only a harness and a verifier, builds the model and recursively improves it against the verifier. The cost is sustained agent API spend plus many local training and evaluation runs on the single GPU, together with the engineering effort to construct a verifier and harness that genuinely reflect the production objective. It can fail by overfitting or gaming the verifier, leaking benchmark information through its own search, emitting non-reproducible artifacts, or running unbounded improvement loops that exhaust API and compute budget without off-benchmark gains.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.18845", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Agent-built models surpass the state of the art on four optical, SAR and multispectral benchmarks and retain most of that accuracy using about 1% of the training labels.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.18845", "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": "" }