FloodSentinel Replay Benchmark

Frozen gauge, warning, and reference corpora for historical U.S. county flood events, with the results of every acquisition policy evaluated on them. The data support the manuscript FloodSentinel: An Agentic AI Framework for Active Information Acquisition and Evidence Governance in Flood Monitoring (Xiao, Murugan, Brody, and Mostafavi; submitted to Computers, Environment and Urban Systems, 2026). Code: https://github.com/YimingXiao98/floodsentinel.

Each county-event pairs 15-minute river-gauge observations with flood thresholds and the National Weather Service warning products issued for that county, each stamped with the time it became available. A monitoring policy can therefore be replayed against the past without lookahead and graded on when it would have observed flooding.

Contents

Path	Contents
corpus/study/	Study 1: Hurricane Harvey (Harris County, TX) and Hurricane Irma (four Florida counties), with the Duval County stage-datum calibration and the Harris County news-corpus manifest
corpus/study_v3/	Study 2: nine held-out county-events and two coverage-only cases, their warning streams, the selection artifacts, the development grid, and the candidates screened during selection
corpus/study_v4/	Study 3: 24 held-out county-events and four coverage-only cases, their warning streams, the selection artifacts, and the candidates screened during selection
corpus/study_v5/	News-verification extension: merged official and news trigger streams for six county-events
results/	Source-of-record results for every study, the post-hoc analyses, and the deployment audit summaries
protocols/	The frozen study protocols, whose hashes the result manifests record
storm_events/	The nine NOAA Storm Events "details" bulk files (2016 to 2024) used for the independent comparison, archived because NOAA replaces them when it revises the database

Study numbering follows the manuscript; the directory names follow the development history (Study 1 is study, Study 2 study_v3, Study 3 study_v4).

	Study 1	Study 2	Study 3
Episode-evaluation county-events (coverage-only)	4 (1)	9 (2)	24 (4)
Reference flood episodes	6	10	46
Gauge observations	28,866 (1,244)	105,017 (20,891)	211,199 (36,142)
Qualifying warning events	n/a	99 (1)	196 (4)

File formats

All corpus and result files are JSON.

Gauge corpus (one per county-event, for example corpus/study_v4/acadia_22001_hurricane_delta_october_2020.json): county, state, FIPS code, the analysis window, and a list of sites. Each site gives its USGS site number, NWS location identifier, name, the basis for assigning it to the county, its action, flood, moderate, and major stages with their source, and a list of observations with fields t (UTC), stage (feet), flow, and q (USGS qualification codes as archived; codes such as A, approved, were assigned after review and were not known at observation time).

Warning stream (*_vtec_triggers.json): the county and forecast-zone codes, the archive source, and a list of events, one per deduplicated VTEC event identity, with phenomenon and significance (FF, FA, or FL warnings), the issuing office, the product identifier, available_at (transmission time, the only time a replay may use to expose the product), begins_at (the published hazard begin time), and audit_final_expire (retained for audit only; it embeds later cancellations and must not be used as an input). The file's as_of_semantics field states these rules.

Manifest (manifest.json in each study directory): every county-event's window, analysis role, source event name, and SHA-256 hashes binding each corpus and warning stream to the results. Selection artifacts (holdout_selection.json, selection_provenance_addendum.json): the event frame, every candidate considered, exclusion reasons, and the rules, applied without reference to policy outcomes.

Results: results/study_v4_additive/counties/*/result.json holds, for each Study 3 county-event, every policy and every clock offset's acquisitions, observed episodes, unobserved flood duration, phase durations, and termination (schedule_rows), and each reference episode's start, end, and warning coverage (trigger_accounting). The other directories hold the corresponding Study 1 and 2 results and the post-hoc analyses reported in the manuscript.

How the reference is defined

At each 15-minute step, a county is flood-positive when at least one fresh gauge with a compatible flood threshold is at or above flood stage, negative when all reference gauges are fresh and below it, and unknown otherwise (1 h maximum age). It grades when a schedule observes flood stage at the same gauges a policy can see; it is not a measure of inundation or impacts. The storm_events/ files allow an independent comparison with report-based records.

Sources and license

Gauge observations and thresholds originate from the U.S. Geological Survey (NWIS instantaneous values) and NOAA's National Water Prediction Service; warning products are National Weather Service products retrieved from the Iowa Environmental Mesonet VTEC archive; event selection used NOAA NCEI's billion-dollar disaster register; Storm Events files are from NOAA NCEI. The derived corpora and results are released under CC BY 4.0; please also acknowledge the original sources. The GDELT news corpora used in the news experiments are not included; they regenerate exactly from the GDELT archive with the script in the code repository.

Limitations

The county-events are U.S. counties, most affected by tropical cyclones, with gauges that carry compatible flood thresholds. Tidal-datum thresholds were excluded. Archived observations may include revisions made after the event.

Citation

A citation will be added when the manuscript or its preprint is available.