Turn stream measurements into a defensible human review queue.
For monitoring coordinators and research QA reviewers: identify
which records deserve attention, see the evidence behind every flag, and record
a human disposition. The model prioritises; the reviewer decides.
Synthetic demonstration: every site, indicator value and result below is invented. No OneAquaHealth measurements are included.
This percentage sets the target share of observations inside an interval. It is not prediction accuracy, a probability that a record is correct, or a water-safety score. Wider intervals can leave more records unassessable.
Pathogen indicator against faecal indicator
consistentoutside envelope — re-checkenvelope too wide to judge
The plot is keyboard reachable: Tab into it, then
←→ to move between sites,
Home / End for the extremes, and f to jump
to the next queued site. ARIA on SVG shapes is unevenly supported, so the
same data is also below as an ordinary table — that is the path that works
everywhere.
Every site as a tableInspect synthetic results, method and limitations
How the example works
The five fictional city groups are held out one at a time. A ridge model predicts one invented indicator from another; a separate calibration split sets the conformal envelope. A reading outside that envelope enters the human review queue. All figures below are calculated from the synthetic fixture.
Illustrative operating characteristics
Observed synthetic coverage is the fraction of invented observations inside their intervals, including intervals too wide to assess. It is not accuracy or a safety guarantee. Current observed coverage: —.
Synthetic coverage and review workload at each confidence level
target coverage
observed synthetic coverage
width
of range
queued
These held-out results describe only the invented fixture. They cannot establish real-world calibration, cross-city transfer or operational impact.
Deployment path
nowOffline proof of conceptStatic, reproducible and usable without network access.
nextCity-specific calibrationUse local history when a city has enough repeated observations.
thenMonitoring integrationConnect intake, reviewer identity and an authenticated audit trail after governance review.
The envelope is too wide to judge these records. They are not automatically cleared or classified as errors. Choose a follow-up or defer with a reason.
Review handoff
Exports contain saved decisions at their original operating points, including reasons and synthetic evidence. This editable handoff does not assign an owner or recipient, track deadlines or delivery, or collect re-test results. It is not an authenticated audit trail. Saving or exporting does not send a re-check request to anyone.