QL-2026-01 Coverage of evidence, Google Flood Hub: pre-registered protocol v1.9

Abstract

Pre-registered protocol for an independent evaluation asking how many people live in reach of a river forecast whose accuracy has never been published. The study reanalyses the developer's own released per-gauge metrics, weighted by the population exposed behind each forecast point, and stratified by settlement mapping density so the answer can be read separately for well mapped and thinly mapped ground. This study makes no claim about the model's accuracy anywhere. The absence of a published metric is not evidence of poor skill, and any reader who comes away thinking that unmeasured means unreliable has been misled rather than informed. Status at deposit: no metric value has been read. The protocol, the basin frame of 230,028 African HydroBASINS level-12 units, the population weights, and all five strata that do not require a metric value were built blind, before the outcome was reachable. The analysis has not been run. Protocol v1.9, sha256 b05c0b932ddb9c6db2f192ff5804570197470175e50ed21ce7520839321d2248. Amendments are appended to section 11 and the frozen text above it is never edited, so every version remains verifiable at its own git tag, v1.0 through v1.9. Version 1.0 hashes to 52ed62354fe58577c77661de5d20dc92a2991db84ad126d92c375e245efb71da. The deposit contains the protocol, its hash file, its OpenTimestamps proof, and a full archive of the repository, which carries the append-only decisions log and run log. Those logs record judgement calls, bugs, and deviations as they happened, including a self-reported period during which the study's own structural-blinding guard was inert, and a population-surface disagreement of roughly three and a half times in sparsely mapped basins that constrains how the results may be reported. Input data is not included, because it is not ours to redistribute and runs to hundreds of megabytes. Every input is re-fetchable from the URL and sha256 recorded in the manifest inside the archive. New in v1.9, all four decided blind. Gate F's comparator is specified as rung 2 of the denominator ladder, the 3,090 real African gauges, after a dry run against permuted labels showed the gate's verdict flipping across all three rungs on data with no real association. F1 is recorded as derived by us from the released precision and recall rather than published per gauge. The mapping-density tercile table must carry its structural floor in the same table, because the gradient is largely a statement about which forecast points hold a metric file at all. And the Note must report the floor of the primary metric alongside it: from file presence alone, 55,234,096 of the 71,248,661 people in reach of an African forecast point are in reach of no gauge holding a metric file, so the primary metric is confined to the range 77.5 to 100 percent before any value is read.