Methodology

The basin snapshot

Most pages draw on one snapshot per water division: river flows, active calls, reservoir storage, snowpack, drought and weather. The Worker fetches each section from its source and caches the snapshot for about half an hour. Every figure carries its source and the date or time the source stamped on it. If a source fails, its section is left empty and labeled unavailable; BasinMind never falls back to a number typed into a page.

How often each source is read

Sources have their own rhythm. Reservoir telemetry posts the next morning, so yesterday's storage is current. USGS posts flow every 15 minutes, so yesterday's flow is not. Drought Monitor maps are released on Thursdays and are valid for the previous Tuesday.

How stale data is labeled

The status pill in the page header reads:

• Live — flows and calls are fresh, and at most one other section is stale;
• Partial — flows or calls are stale, or two or more other sections are; hovering the pill names them;
• Offline — nothing could be read.

When a refresh fails, the previous figures are carried forward with their original date and judged by their own age, so a reading from an hour ago stays fresh until its window passes. Individual figures show their own “as of” time.

Percent of median

For USGS gauges, BasinMind compares the latest instantaneous flow with the USGS daily median for the same calendar day over the gauge's period of record, and rounds the ratio to a whole percent. Flows are described as above normal at 110% or more, normal from 75% to 110%, below normal from 50% to 75%, and low below 50%. DWR telemetry stations, which fill in the Rio Grande basin, have no such median, so their flows are shown without a percentage rather than estimated. Snowpack percentages use NRCS's 1991–2020 medians for the date.

How AI answers are grounded

Ask BasinMind answers with a language model that must use tools — functions that read live public data from DWR, USGS, NRCS, NOAA, Reclamation, the Drought Monitor, EPA and others — before answering. Its instructions are to answer first, give every number its source and date, say when confidence is low, and never invent figures. Then two checks run on every answer:

• Figures. Every number in the answer is compared with the numbers in the tool results from that turn (allowing for rounding). Dates, years, IDs and small counts are skipped. When a figure is not found, the in-depth tier rewrites the answer once from the tool results; any figure still unsupported is noted under the answer.
• Missing data is never a “no.” BasinMind's tools cover Colorado hydrology and water rights. If an answer says something does not exist or never happened, on a subject no tool returned data about, it is rewritten to say BasinMind's sources do not cover the topic and where to check. If the claim survives, a note under the answer says that a “no” there means “not found in our sources,” not “confirmed.”

Answers that do not name their sources get the list of agencies whose tools returned data added at the end. Questions that need comparison or analysis go to a stronger model; simple lookups use a smaller one. Your thumbs on answers go to a review queue, and a thumbs-down can become a test case in BasinMind's evaluations.

Known limits

• Advisory only. BasinMind reads public records. Colorado DWR administers water; water courts decide rights.
• Coverage varies by division. Each division is represented by a set of gauges, reservoirs and SNOTEL stations, not every one. Division 3 relies largely on DWR telemetry, so percent-of-median is often blank there.
• Provisional data. Real-time USGS and DWR values are provisional and can be revised.
• Pilot. Sources fail, and AI answers can be wrong despite the checks. Use the thumbs to tell us.