How we reach every number.

The most common failure in property data products is presenting an estimate with the same visual authority as a measurement. We refuse to do that, and the refusal is enforced in code: a value cannot be rendered in a report without declaring its provenance.

• RecordedTaken verbatim from a government dataset. A well log that says 176 feet is a recorded value.
• CalculatedArithmetic we performed on recorded values. A median depth across 37 nearby wells is calculated.
• EstimatedOur model's inference. An expected drilling range for a property with no well on it is estimated. It is not a measurement and never presented as one.
• Potential matchA spatial or probabilistic association rather than an authoritative link. Every water-right match is a potential match.

A depth estimate is only as good as the wells it draws on, so selection is the most consequential step. We start from every well record within five miles and then exclude records that would distort the picture:

• Sites that are not wells, such as springs and surface-water sites.
• Wells with no recorded total depth.
• Wells recorded as abandoned, plugged, destroyed, or dry — they describe holes no longer in service.
• Depths outside a plausible range (under 5 feet or over 2,000 feet), which indicate data entry errors rather than unusual wells.
• High-capacity production wells — municipal, public supply, irrigation, industrial, mining, injection, and dewatering. These are drilled to different targets and at different diameters than a house well. This matters in practice: near Montana State University we found a 500 GPM research well a few hundred metres from ordinary 20 GPM domestic wells.

Your report lists every exclusion and its count, so the comparable set is auditable rather than a black box.

We then expand the search radius progressively — 0.5, 1, 2, 5 miles — and stop at the first radius that yields at least 10 comparable wells. We report the radius used. If five miles still yields fewer than 5 comparables, we decline to publish a range at all and say so.

Geology outranks proximity. A well slightly further away in the same geologic unit is a better analogue than a closer well across a boundary. If enough same-geology wells exist at a given radius, we use only those.

Comparables are not counted equally. Each receives a weight that is the product of five factors, and the report shows the resulting statistics as weighted values:

• Distance. Inverse-square decay from the subject property, normalized to the radius in use.
• Geologic similarity. Graded rather than binary. An exact formation match counts fully; the same geologic system counts at 0.8; merely similar rock character counts at 0.4; materially different rock is discounted to 0.18. Cross-formation extrapolation is the least defensible inference available, so it is discounted steeply.
• Elevation difference. Depth to water tracks topography closely in Montana's intermontane basins. Wells within 50 feet of the subject elevation count fully; several hundred feet of relief usually means a different hydrologic position.
• Record age. Recent logs are preferred, but older logs still inform. Undated logs are downweighted rather than discarded.
• Location accuracy. Well coordinates derived from a legal land description rather than measured can be off by a quarter-quarter section — roughly 1,300 feet. Those records are downweighted.

An honest note on geology coverage. Montana's Ground Water Information Center assigns an explicit aquifer designation to only about a quarter of wells. Where a well has no designation, we determine its geology by locating it in the state's aquifer polygon layer — the same method used for the subject property, so the comparison stays like-for-like. But that layer is coarse, so when geology is established this way we cap the geologic component of the confidence rating rather than treating a coarse match as strong evidence. Your report states which basis was used and for how many wells.

We never publish a single predicted depth. A single number implies precision that comparable-well data cannot support. Instead every report gives two bands:

• Estimated likely drilling depth — the weighted interquartile range (25th to 75th percentile) of comparable depths.
• The wider context band — the weighted 10th to 90th percentile, stated as "80% of comparable nearby wells fall between X and Y feet", so tail risk is visible rather than hidden.

Rounding is tied to confidence. A high-confidence estimate rounds to 5 feet, moderate to 10, low to 25. Reporting "163.7 feet" from fourteen comparables would be false precision, so we do not.

This is a transparent statistical model, not machine learning, and we do not claim otherwise. With a few dozen relevant wells per site, a well-specified weighted percentile is both more defensible and more explainable than a fitted model.

Five components are assessed and all five are shown to you:

• Comparable well count.
• Geologic similarity, and how that geology was determined.
• The search radius that was needed.
• Depth consistency — the interquartile spread relative to the median.
• Record recency.

Confidence is deliberately conservative. Fewer than 5 comparables caps the rating at LOW no matter how the other components score, because no amount of methodological care rescues a sample that small.

Yield uses the same comparable set, restricted to wells with a recorded yield. We report the weighted median, the typical range, the distribution, and the share of nearby wells below thresholds that matter — including 35 GPM, which is Montana's exempt-well flow limit.

What a recorded yield is. It is what a driller measured during a short test at completion. In the data we have loaded, many of these tests ran one or two hours, and the report tells you the actual test durations behind your numbers. A drilling-test yield is not a guarantee of sustained production; long-term capacity depends on aquifer properties, seasonal recharge, and pumping by others.

No Montana dataset authoritatively links a parcel to a water right. We therefore match spatially and label every result as a potential association with its basis stated:

• Likely associated. A recorded point of diversion falls inside the parcel boundary. The strongest signal public data offers.
• Possible association. A diversion sits immediately adjacent to the parcel, or the parcel falls inside a place-of-use tract. DNRC maps places of use by legal land description, and we verified that these tracts commonly run about ten acres and up — usually larger than a single parcel. So a place-of-use match may well belong to a neighbour.
• Context only. A diversion in the surrounding area with no indication it relates to the parcel.

We will never state that a property does or does not hold a valid water right. That is a title question requiring the original records and, usually, a qualified professional.

The primary output is five component ratings — well depth, yield outlook, water rights, regulatory, and data confidence — because a single number hides exactly the nuance a buyer needs.

A composite Water Outlook is shown only when it can be trusted. Two rules govern it:

• When data confidence is LOW, the composite is suppressed entirely. A confident-looking score resting on thin data is worse than no score.
• When a red flag is present, the composite is capped below the GOOD band. A property inside a closed basin, where a new well may not be permittable, must never average out to "GOOD" on the strength of good depth and yield figures. The regulatory question is a gate, not one input among five.

• We do not fill gaps with generated text. If a record is missing, the report says the record is missing.
• We do not state a groundwater trend unless the monitoring record genuinely supports one. Inferring a trend from well completion dates would be misleading.
• We do not tell you whether water is safe to drink. Nearby sampling is area context; only a laboratory test of the actual well answers that.
• We do not claim to know whether a property lies in a stream depletion zone. Montana applies a stricter exempt-well threshold inside one, and no authoritative public map of those boundaries exists, so we show both thresholds and point you to DNRC.
• We do not present a database match as a legal determination of anything.