CZDTS reads the whole satellite record held for an asset, not the latest scene, and issues a dated statement of what it measured, with the uncertainty carried through, and a plain list of what it could not measure. It runs on free data by default, buys a capture only when a specific asset raises a specific question, and takes your own instrumentation into the same record.
One site, as the report renders it: eight tailings facilities at a working copper mine in Arizona, watched for 3.7 years.
Draw an area and the platform pulls the archives it has indexed for it, then keeps appending, and the statement prints which archives it read. At the Arizona site that came to 22,108 state rows across eight facilities in 3.7 years. Normal is defined by each facility’s own history, so the detector needs no failures to run.
Coverage differs by site and the statement prints the window it read. The earliest data anywhere in the platform is June 2018; the median area begins in 2024. An area with no usable imagery is reported as unobserved. We do not report it as unchanged.
A self-referenced check says a facility has changed. It does not say it was ever safe. A facility that has been degrading slowly since before its record begins carries that degradation inside its own baseline. So every invariant is marked in the statement as learned from history or taken from a design value, and where you supply a site plan the check runs against the design value instead.
Each kind of asset carries its own list. For a tailings impoundment: the crest displacement rate in line of sight is not increasing, the pond is not growing beyond this facility’s own seasonal range, the minimum beach width is not reducing, no new wet area appears on the downstream face, and no new or expanding seepage appears at the toe beyond what the drains discharge by design. Each is tested separately against that facility’s own record, and the statement names the ones that could not be tested.
Every check returns one of four states. The last two are different problems with different answers.
The check ran on this asset in this window and the state sits inside what the asset’s own history predicts. This is the only state in which an empty report means something.
Outside expectation past a stated threshold, held for a stated length of time. You get the alarm, the before and after, the series, and the run it came from.
The archive could not support the question this time: no scene, an incoherent interferogram network, no named control. A later pass or a tasked capture can fix it. No claim is made meanwhile.
Outside the method at any resolution: the phreatic surface, pore pressure, internal erosion, drain condition, pond depth and stored volume, and the state of the tailings themselves. No satellite capture will answer these; an instrument will.
The distinction between the last two is the one that decides what to do next. The first is a gap you can close by spending money. The second is a gap you close by installing an instrument. Freeboard as an elevation sits between the two: it is not in the free layer, but a tasked stereo capture can return it, and the statement says which was used.
A detection on its own cannot be underwritten, because nothing in it says how often the system is wrong. So each statement carries the state with its uncertainty, the alarm count with the exposure it was measured over, and the share of checks that could not be run. Uncertainty is propagated from each input stream; calibration against held-out ground truth is not yet demonstrated and is not claimed.
At the Arizona site the alarm rate is 0.14 per facility-year over 29.6 facility-years, with a 95% interval from 0.04 to 0.35. All four alarms fell in one winter window across different facilities, so they may be one weather event counted four times. We publish the interval because the point estimate on four events is not a precise number.
At a site with no recorded failure the statement reports detectability against injected onsets, and labels it as such. No lead time has yet been measured against a real failure.
The free layer is wide, continuous and coarse, and it carries the screening. When it raises something, a sharper capture or an independent one is a single order away. Where the sky cannot settle it at all, your own instrumentation joins the same record.
Choosing a capture means knowing who images the wavelengths in question, what they publish as a rate, and what minimum order they will bill you for. We keep that survey current and public: the Earth observation procurement reference covers the operating providers we have found across optical, radar, hyperspectral and thermal, alongside the free public missions and the zero-cost routes open to researchers, and it names the operators who publish no rates at all.
Customer observations are ingested against the same asset and the same run lineage, and are used for calibration and validation. Combining them with satellite evidence inside the state estimate is the open research problem, not a build task: no forward model of how facility state evolves transfers between sites, and when streams disagree those regimes carry no labels.
Lenders, auditors and their technical advisers buy a single dated artefact on a named asset. The rows below that one are who else it serves.
| Buyer | What they buy |
|---|---|
| Lenders, auditors and their advisers |
A dated verification statement on one named asset. It says what was measured, over what window, with what uncertainty, and lists what could not be measured and why. Your adviser can reproduce any trigger in it from the API alone. Facility outlines are proposed from imagery unless you supply a site plan, and the statement says which it used. On a covenant date where a check comes back unobserved, the statement says so rather than reporting the asset as unchanged. |
| Operators and Engineers of Record |
An independent read alongside instrumentation, and the evidence trail conformance work needs. A statement issued against an outline proposed from imagery is a screening product and is labelled as one. Supply a surveyed outline and the statement records that it used it. |
| Registries and regulators |
A check on a declared value that does not depend on the declarer, with the observability limit stated beside every number. |
| Insurers | Not yet. A parametric trigger needs a frozen index definition, a stated rule for what happens when a displacement epoch is unobserved, and a lead time measured against real events. We have the first, are working on the second, and do not have the third. |
Thirty-three invariants are declared in the specification, twenty-nine are implemented in code, and eight have been measured on a real asset. The engine can run the rest. It has not.
| Asset family | Examples | Measured on a real site |
|---|---|---|
| Impoundment | Tailings dams, ash ponds, evaporation and leach ponds | 8 of 8 |
| Conveyance | Canals, pipelines, levees, diversion channels | 0 of 5 |
| Field surface | Irrigated land, plantations, managed rangeland | 0 of 5 |
| Closure | Capped landfills, closed mine sites, remediated ground | 0 of 4 |
| Earthworks | Embankments, cuttings, stockpiles, haul roads | 0 of 4 |
| Managed system | Wellfields, water networks, subsurface programmes | 0 of 4 |
| Terminal and plant | Refineries, tank farms, concentrators, data centre campuses | declared only |
A methane and emissions overlay is declared once and binds into subsurface risk and closure, rather than being an asset family of its own. Measured means a rule with a stated threshold produced a result on a real asset from real data.
Each statement records the data it used, the code version and the settings it ran under. Your own analyst can pull the same number back from the API and get the same value, whether or not it raised an alarm, for as long as your report is live.
Corrections are issued as new versions with their own identifier, and the original stays readable, so a figure quoted in a meeting last quarter can still be looked up.
Licence terms follow the data a run actually consumed rather than a label attached by hand, so a product inherits the most restrictive licence among its inputs, and where an input cannot be redistributed the statement says so.
Give us an area and the sentence you need to be able to write about it. We will come back with what the free archive actually holds for that ground and from when, which checks would run, which would come back unobserved, and which parts of your question are outside the method entirely. If the answer is that we cannot help, we will say so.
west, south, east, north, a GeoJSON polygon, or a WKT string. If the outline came from a
site plan rather than from imagery, say so, because it changes what we can put our name to.
Prefer to write it yourself? Email us directly. CZDTS LLC, Tempe, Arizona.