Source-linked AI summary
CDEP Agent: Connecting Meteorologically Detected Temporal Compound Events to Real-World Documentary Evidence
Zhuoran Li, Weiyi Kong, Boer Zhang
TL;DR
The study asks whether drought-to-extreme-precipitation transitions detected meteorologically also enter official records and public webpages. CDEP Agent links fixed California candidates to provenance-retaining evidence and finds limited documentation of both hazard components and their compound linkage.
Problem
Whether meteorologically real CDEP transitions appear in real-world warning and post-event records remains insufficiently documented.
Method
CDEP Agent retrieves and aligns official records and public webpages with fixed meteorological candidates using explicit evidence dimensions and auditable matching checks.
Results
34.3% of 408 California candidates were corroborated on both hazard components, and 1.5% were explicitly linked to their antecedent drought.
Takeaways & Limitations
The provenance-linked records provide a basis for comparing meteorological event definitions with what official and public records document.
Takeaways & Limitations
The study covers California during 2021–2025 and uses one operational CDEP definition, so support rates and regression estimates may differ elsewhere or over longer periods.
Abstract
from arXiv · showhide
Compound drought-to-extreme-precipitation (CDEP) events are recognized in climate science as a growing driver of extreme impact, but whether this recognition carries over into real-world early warning and post-event documentation is unknown, so a meteorologically real CDEP event may pass with neither advance warning nor any later record. Here we present CDEP Agent, an auditable LLM-agent framework that tests this mismatch directly by linking CDEP candidates detected from meteorological reanalysis to real-world hazard and impact evidence across sources with different spatial scales, temporal resolutions, and reporting conventions. Using California as a case study, we identify 408 candidate CDEP events from ERA5 observations during 2021-2025 and evaluate each against the U.S. Drought Monitor, NOAA Storm Events, and public webpages along five dimensions: antecedent drought, extreme rainfall, local impact, hazard-impact attribution, and explicit drought-to-rainfall linkage. Only 34.3% of candidates are corroborated on both hazard components, and just 1.5% are ever explicitly linked to their antecedent drought, indicating that most meteorologically detected CDEP events go undocumented and their compound nature almost never enters the record at all. Our framework gives climate scientists a way to test physical event definitions against what actually gets documented, and gives social scientists, economists, and disaster-response agencies a provenance-linked evidence base for compound events that current warning and reporting systems largely fail to capture.
Introduction
Meteorologically real drought-to-extreme-precipitation transitions may not appear in the records used for warning and disaster response. CDEP Agent tests this documentation gap by linking fixed meteorological candidates to official records and public evidence.
- Introduction: Compound events can be meteorologically real yet poorly documented because records are organized around separate storms, droughts, and jurisdictions.
- Introduction: CDEP Agent connects meteorologically detected candidates to hazard and impact evidence while retaining provenance and supporting quotations.
- Introduction: The framework evaluates antecedent drought, extreme rainfall, local impact, hazard attribution, and explicit drought-to-rainfall linkage.
- Introduction: California provides a consequential case study, including the Oroville Dam spillway failure and evacuation of roughly 190,000 downstream residents after drought-ending rainfall.
- Introduction: 34.3% of 408 candidates were corroborated on both hazard components, while 1.5% were explicitly connected to their antecedent drought.
Methods
The study constructs fixed California CDEP candidates from ERA5-Land and then applies candidate-specific retrieval, structured review, and deterministic matching checks. Evidence is retained only when dates, locations, hazards, consequences, and event relationships satisfy the stated requirements.
- Constructing Historical CDEP Candidates from Meteorological Data: Candidate construction used SPEI-3, a location-specific extreme-precipitation threshold, and a maximum three-month transition interval.
- Retrieving and Reviewing Evidence: For each fixed candidate, the agent matched U.S. Drought Monitor and NOAA Storm Events records before searching public webpages for specified hazards and impacts.
- Constructing Historical CDEP Candidates from Meteorological Data: Extreme precipitation was detected from three-day accumulated rainfall using each grid cell’s P99 threshold from the 1985–2014 reference period.
- Retrieving and Reviewing Evidence: Follow-up searches targeted missing drought, extreme-precipitation, impact, or linkage evidence while preserving candidate location, relevant time period, and evidence type.
- Retrieving and Reviewing Evidence: Underlying page text was required for review, and inaccessible sources were recorded as limitations rather than treated as evidence of absence.
- Assigning the Five Decisions: A No decision meant that completed review found no qualifying record under the rules, not that the event or impact was absent in reality.
- Human Validation of Retrieved Records: Manual review agreed with 78 of 80 event assessments and found two additional extreme-precipitation records missed during retrieval.
Results
Across 408 California CDEP candidates identified during 2021–2025, documentary support was substantially stronger for antecedent drought than subsequent extreme precipitation, and qualifying webpages were more common for longer precipitation events. Public webpages that passed review usually documented local consequences and linked them to rainfall or flooding.
- 1.03 ± 0.52 percentage points per year was the linear trend in the proportion of California grid cells affected by at least one CDEP event during 2000–2025.
- 34.3% of 408 candidates were corroborated on both the antecedent drought and subsequent extreme-precipitation components.Records supported drought in 72.8% and extreme precipitation in 45.8% of candidates.
- 1.40 times the odds of finding a qualifying webpage was associated with each additional day of extreme-precipitation event duration.The 95% CI was 1.04–1.88, with p = 0.025; the estimate remained similar in sensitivity analyses.
- 41.9 days versus 49.1 days was the average drought-to-extreme-precipitation interval for candidates with versus without qualifying webpages.The adjusted odds ratio was 0.72 for each additional 30 days, while relationships for drought extremity and antecedent drought duration were less clear.
- Documentary corroboration did not represent meteorologically detected candidates equally, potentially reflecting event salience, reporting practices, source availability, or retrieval coverage.
- 40 of 45 qualifying webpage records documented at least one local impact, and 37 of 45 explicitly connected that impact to rainfall or flooding.Reported consequences included flooding, road disruption, outages, evacuations, debris flows, landslides, property damage, and rescue or medical response.
- Qualifying webpages were found in every year from 2021 through 2025, across 26 storm families and 26 counties.Accepted documentary evidence occurred across major urban counties and less-populated inland and rural counties.
Discussion and Conclusion
CDEP Agent links meteorologically constructed candidates to structured records and public webpages through a candidate-specific, provenance-retaining evidence review. The study shows that hazard components may be documented separately while their compound relationship remains absent, but its estimates are bounded by California’s 2021–2025 scope and one operational definition.
- The review connected each meteorological candidate to antecedent drought, subsequent extreme precipitation, and documented local consequences while retaining sources, quotations, dates, locations, and matching results.
- CDEP Agent retains source passages, matching results, and decision rules to produce auditable assessments rather than untraceable model judgments.
- Each additional day of extreme-precipitation duration was associated with 1.40 times the odds of finding a matching public webpage, while other candidate-characteristic relationships were less clear.
- 40 of 45 qualifying webpage records documented local consequences and 37 explicitly connected those consequences to rainfall or flooding.Structured records helped confirm hazards, while public webpages often supplied local consequences affecting residents and infrastructure.
- Preserving both hazard components and their relationship can support future datasets designed specifically for sequential compound climate events.The procedure can be applied to candidate sets defined with different thresholds, transition windows, regions, or event types.
- The study evaluated California during 2021–2025 using one operational definition based on SPEI-3, a location-specific precipitation threshold, and a three-month transition window.Support rates and regression estimates should not be assumed to remain unchanged in other regions or longer periods.
Supplementary Material
The supplementary material details an auditable retrieval-and-review workflow that matches documentary evidence to candidate events under explicit checks, then reports decision distributions and sensitivity analyses.
- Retrieval and review workflow: Follow-up searches target only unsupported evidence categories, including antecedent drought, extreme precipitation, local impact, and drought-to-extreme-precipitation linkage.The query-generation model receives the candidate location, dates, current evidence target, and already accepted evidence; generated queries must pass location and period checks.
- Retrieval and review workflow: Each candidate record combines official records, generated queries, retrieved URLs, page reviews, accepted and unresolved evidence, visited pages, and remaining retrieval budget.The running record lets follow-up searches depend on evidence already obtained while preserving the actions leading to final assessments.
- Page-level evidence validation: Only clearly matched webpage records can support a decision, with date overlap, county or within-county location, observed-hazard, same-event, quotation, and candidate-specific evidence checks enforced.Hazards, consequences, and stated relationships from different storms are not combined, and co-occurrence on one page is insufficient without an explicit connection.
- Sensitivity analyses: 1.37 adjusted odds ratio for extreme-precipitation event duration remains significant after excluding 68 candidates, while clustered sensitivity analysis yields 1.40.The exclusion analysis includes 340 candidates; the adjusted odds ratio is 1.37 (95% CI 1.02–1.85; p = 0.038), and the clustered estimate is 1.40 (95% CI 1.04–1.88; p = 0.026).