Internal · operator tools
Reviewer tools
The human-in-the-loop surfaces that turn “an automated signal fired” into a verified, dated, cited outcome. These build a labeled-outcome corpus that's hard to copy — the basis of the fusion-calibration advantage. They're operator tools, not a public feature.
Flood event labeling
Real candidate flood signals (GDACS / GDELT / RSS / IMD), each with its source text, a map link, and before/after Sentinel-2 imagery — a human decides “did this flood happen” with evidence in front of them.
Review board
Four columns on the existing queues: event candidates, cell-periods where automated sources disagreed or the band is wide, confirmed verdicts and rejected records. Every verdict carries the reviewer’s name and the time.
Place risk cards (preview)
The per-place flood-risk card for every place slug: this-week probability and band, top reasons, relative risk and depth rows, each a real committed value or “not computable” with its reason. Held here, not on the public place pages, because on every place more than half its fields are not computable.
Why these are first-class
Each model here has a known accuracy ceiling, and we state it — the lasting asset is the labeled-outcome corpus these loops produce, and the loop that fits calibration weights to it. Keeping the reviewers discoverable (not a file path you have to know) is the UI half of that moat. Both tools run against real, already-committed signals.