Colorado articles · Records and method
Where a Colorado report gets its records
A report reads records from many agencies that share no identifier. These are the places where the joining is done.
Every record behind a Water Table report is public. That is not the moat and it was never going to be. Anyone can query the Division of Water Resources. What is genuinely hard is that the answer to "is the water on this listing real" is spread across agencies that do not share an identifier, publish on different schedules, and in several cases do not publish a statewide layer at all.
Not one agency's API. Every feed is joined at the survey section, including county assessors that publish their own parcels and appear in no statewide layer.
Water Table source provenance · computed September 2026
The counties that are not in the statewide layer
Colorado has no single authoritative parcel layer. There is a statewide composite, and it is thin or absent in exactly the counties where rural land actually trades. Costilla County, the county with the largest concentration of cheap subdivided rural parcels in the state, and the one where a buyer is most likely to be sold ground with no water, contributes tens of thousands of parcels that have to be pulled from the county assessor’s own service.
One of the source lines in the provenance table reads "Fremont County Assessor GIS, and Chaffee County published by the City of Salida". That is what this work looks like at the edges: a county whose parcels are served by a municipality inside it, found by looking, not by a directory. Nobody builds that from an API catalog.
The joins neither agency publishes
The Division of Water Resources publishes structures and the rights filed against them. It separately publishes an irrigated-lands mapping: parcels, the crop on them, the irrigation method, and which structures supply each one. Nobody publishes the two joined, and the join is where the useful sentence lives.
The join nothing else publishes. A decree names a structure; the state's irrigated-lands mapping names parcels. Putting them together turns "the J & M Hughes Ditch exists" into "it waters these parcels, this many acres, growing this".
DWR structures against the state irrigated-lands mapping · computed September 2026
That is the shape of most of the work here. Almost nothing in a report comes from a single query against a single agency: it comes from putting a structure beside a parcel, a well beside a scanned log, a court filing beside a survey section, and keeping every one of those joins honest about how confident it is.
Some of the record was never typed up
The state images the driller’s own completion report for wells and the scanned page often carries values the tabular record does not. Those images are cataloged and read page by page by a vision model, and the extracted values are set against the API’s, field by field, with the disagreements kept rather than resolved.
DWR images the driller's own completion report and the tabular record does not always carry what is on it. These are read page by page by a vision model and set against the API's own values, so a disagreement is surfaced rather than averaged away.
DWR imaged well documents, read by a vision model · computed September 2026
The same principle runs through the water court resumes, which are prose rather than data. Every filing carries its case number and a legal description of the ground, written the way a clerk wrote it in 1974 or last month. Turning that into "this filing is about this square mile" is a parsing problem with no specification, and it is tested against the corpus rather than against an assumption about the format.
What we refuse to do
- No averaging a sentinel. Agencies write -9999 and 0 to mean "not recorded". Those are nulled, never treated as measurements, and never clamped into range.
- No arbitrary pick. Where a record could supply several values, the report does not take the first one it finds and print it as the answer.
- No silent gap. A subject with nothing on record is printed as having nothing on record, in words, rather than being left out so the page looks fuller.
- No inference dressed as a record. Every answer states its basis, and an estimate says what it was estimated from.
Those four are not aspirations. Each of them is there because it was got wrong once, found, and fixed, and each has a test that fails if it comes back.
What the water provider charges and allows
No agency publishes what a town or district charges to connect a new home, or what it lets a household water. Each provider posts its own, as a fee schedule, a section of city code, a resolution or a drought page, and many post nothing. Every tap fee and watering rule in a report is read from the provider’s own page by a person and kept with a line quoted from it. Each month the page is fetched again, and a rule whose line has gone is no longer printed.
The season is the trap. Denver Water ended lawn watering on October 1, 2026, and so did most of the districts it supplies, while their summer pages still said two days a week. A rule read from a summer page in October is wrong, so a watering page is read for its date before its rule.
Everything is dated
Every source carries when it was last loaded and how far behind the publisher it is. Some of these move on a decade timescale, and irrigated-lands mapping is redone every few years, while others land weekly. A report that presents a decade-old irrigation map and a Tuesday well permit with the same confidence is lying by omission, so the report prints the age of what it read.
This is general information about Colorado law and records. It is not legal advice or a title search. What applies to a particular parcel is settled by the documents filed against it.