We let the river grade the model.
Some parcels of land quietly hold up the water, the farms and the towns downstream of them. This engine finds those parcels, names who depends on each one, and scores that dependency with honest uncertainty, so nature can be underwritten like the infrastructure it already is. Across km of the Colorado, from the Eagle headwaters to the Moab canyon, we then handed it to the data, testing those scores against real monitoring stations and reporting, in public, that they do not yet beat a simple two-variable baseline. That reckoning is the story.
candidate parcels · catchments across sub-basins · named beneficiaries · one run · live public data
Capital cannot underwrite a public good it cannot point to.
The world faces a biodiversity finance gap on the order of 700 billion dollars a year, and more than half of global GDP already leans on nature. Yet no market can buy “healthier watersheds in general”. It can only underwrite a specific asset tied to a specific cashflow at risk, and that is exactly where today's tools run out of road. They can score a site's ecological condition, but almost none can tell you who downstream actually depends on it, or by how much. So capital stalls, not for want of money but for want of an address, right at the moment of pricing.
$700B / yr biodiversity finance gap · more than half of global GDP is nature-dependent · biodiversity credits are still priced cost-plus, with baselines seldom validated. These are market-wide figures, not this corridor's own value, which we have not tried to price.
An ecosystem is a network. Its value has to be priced like one.
Species counts are to ecosystem health what a headcount is to a company, a tidy number that flatters and hides the thing that actually carries the risk. Value in a living system does not add up site by site; it moves along edges. A keystone is the node a whole neighbourhood of the network leans on. A spillover is what travels down those edges. A beneficiary is where that flow finally lands as a cashflow. Resolve the dependency graph, from source through conveyance to demand, and for the first time you can say which parcel underwrites which payor. That graph is what turns something ecology can only watch into something a market can price. It is the engine.
The keystone polygon: the unit finance was missing.
In any living network, the pieces that matter most are rarely the biggest ones. Ecology learned this the hard way in a tide pool. Remove one unremarkable starfish and the whole community comes apart, its hold on the system wildly out of proportion to its numbers (Paine, 1966). Importance was never about mass; it was about where a thing sits in the web of who depends on whom. The half-century since carried that idea from a single species to a place, to the discrete features a landscape leans on (Tews, 2004), and to the flows of water, sediment and living things that ignore property lines on their way between them (Loreau, 2003). Draw those parcels as nodes and the flows between them as edges, and the word keystone stops being a metaphor and becomes something you can measure, the high-centrality polygons whose loss would fragment the whole network (Urban and Keitt, 2001).
A keystone polygon is that outsized parcel, named and scored. It is the land whose position and pathways hold up the productivity of the estate around it and the people downstream who depend on it, the smallest thing this engine underwrites, one polygon at a time, never a portfolio average and never a site rating.
Foundations: Paine 1966 · Tews et al. 2004 · Loreau et al. 2003 · Urban & Keitt 2001 · Bagstad et al. 2013. Naming the land polygon as the underwriting unit is the new synthesis, set out in The Land Next Door.
One river, three jobs: generate, convey, consume.
We read the corridor as one network that runs from source to conveyance to demand. The headwaters, at 9,000 to 12,000 ft, make the water; the mainstem, at 5,500 to 7,500 ft, carries it; the valley and canyon, at 4,500 to 5,500 ft, spend it. What a parcel is worth then comes down to where it sits in that flow and who lies below it, resolved across NHDPlus catchments (the national stream-and-catchment map) and drawn here over the real -river network from USGS NHDPlus HR.
16 named beneficiaries, routed by hydrology, not by proximity.
From transmountain diversions to Grand Valley farms and Moab's hospital and cleanup sites, every beneficiary is routed to parcels through zone-weighted service profiles. The configuration even bakes in its own corrections. Homestake feeds Aurora and Colorado Springs, not Denver. Ute Water draws first on Plateau Creek, with the Colorado only as a backup. Moab Regional runs on municipal wells rather than a direct river intake. The framework is candid about where it might be wrong.
Every input is public, named, and auditable.
Every input is public government data. One federal service turns any point on the map into its patch of watershed and follows the river up and downstream from it. On top of that, the standard public records for water quality, land and forest cover, soils, fire fuel, protected land and wildlife build the nine ecological measures each parcel is scored on. Nothing here is proprietary and nothing is hidden, so a diligence team can pull every source and check the work itself. For this first run those inputs are modeled stand-ins rather than live per-watershed pulls, and the ledger below says so plainly.
Bayesian keystone screening, humble by construction.
Seven ecological signals, each scored from zero to one, combine into a single reading of how load-bearing a parcel is. What matters is that every parcel comes back not as one number but as a number with an error bar, a best estimate plus the range the evidence actually supports. The model starts by assuming very little and lets that uncertainty flow all the way through, so the width of the range is itself part of the answer. The pass marks are set in advance, keystone at 0.72 and investable at 0.80. The model is built to measure its own doubt rather than hide it.
The bidirectional graph worked, end to end, on one run.
The engine traced forward impact paths, parcel to beneficiary, and reverse dependency paths, matching of 56 parcels in both directions. The load-bearing minority turns out to be small and tightly clustered. Only of 56 (%) carry a posterior whose credible interval clears the prior, and all of them sit in the Eagle headwaters, while a looser screening cutoff admits %. The top of the ranking is .
A real signal, that a two-variable baseline still beats.
Before anyone saw a result, the team sealed the method and the success bar in a hashed, timestamped record, so no test could be quietly bent to fit the answer later, then checked the sediment-avoidance score against real monitoring stations. The signal is real, and catchments that hold back more sediment really do run cleaner. But plain forest cover on its own does better than the full model, the effect nearly disappears once you account for elevation and development, and on watersheds it had never seen, a baseline of just elevation and percent developed ( R squared) beats the whole custom engine. That is why the engine's verdict on itself reads .
It recalibrated its confidence downward.
The model used to borrow its confidence from a made-up reliability number that flattered it. Now its confidence is tied to how well the signal actually held up on data it had never seen, so a mechanism with almost no real skill can no longer pretend to any. Move the control and the engine recomputes a good parcel's estimate live, and you watch it slide back toward its cautious starting point of 0.54. Widening the noise can only pull the estimate toward caution. It cannot manufacture confidence the data does not support.
Proves the engine. Does not yet prove investability lift.
The ledger below is generated straight from the run, not written by hand. The proven column is the machine. The not-yet column is the honest distance to a dollar. The 56 units are candidate envelopes, of them seed envelopes and verified boundaries; no mechanism yet clears validated-incremental, and no cap rate is computed.
This is not modesty for its own sake. The last environmental market taught buyers to distrust unaudited claims. An independent review found that the large majority of a leading standard's rainforest credits delivered no real benefit, and nature-based prices fell from roughly fifteen dollars to about one. The next market will pay for rigor, and an engine that pre-registers its test, checks itself against measured reality and publishes where it fails is built for that world.
Diligence an agent can check for itself.
This ships for the two audiences that now decide where capital goes. One of them is a human
diligence team. The other is the fast-growing population of agents that pre-screen deals on their behalf,
often before a person ever sees them. Every figure here and in the pack carries its own source,
machine-readable and audit-ready, a human narrative sitting alongside an agent-consumable
llms.txt, so an agent can re-pull the data and re-run the check itself instead of quietly
discounting a document it cannot parse. With machine-mediated commerce heading into the trillions by the
end of the decade, and provenance standards consolidating around MCP and llms.txt, an artifact a machine
cannot verify is already invisible to a growing share of capital.
From diligence to distribution.
This engine is the rating desk. It names which polygons are load-bearing, who depends on them, and how sure we are, and it deliberately stops short of printing a price. That last step is the seam a partner fills. BASIN Natural Capital runs the settlement layer, issuing Certificates of Ensurance that finance a place directly and pricing them with a Natural Cap Rate, ecosystem-service value over real-estate value, which turns an ecosystem's worth into a figure a market can hold. Wire the two together and the map stops being a report and becomes a storefront, where every keystone polygon carries a slot for its instrument, ready the moment that place clears monitoring and its instrument is issued.
The rail is already wired to the run's real keys, 56 situs and 16 binder-registered places first in line, with zero instruments issued today. Nothing here is for sale, because the same honesty gate that governs the science governs the storefront, so an instrument appears only where a polygon has earned it. That restraint is the whole point. In a market still healing from carbon's collapse, an honest venue that an allocator and an autonomous agent can both trust, and settle on, is the scarce asset. Diligence becomes distribution, and honesty is what converts.
56 situs · 16 binder-registered places · 0 instruments issued today · settlement via BASIN Certificates of Ensurance
In this market, rigor is the product.
Most vendors would bury “elevation and development beat our model”. This corridor led with it. Every score carries its uncertainty, every claim carries its boundary, every mechanism carries its measured verdict, and the deliverable ships with an artifact-QA gate that fails the pack the moment it asserts validated value for a mechanism that has not earned it. That is the asset, a dependency-graph engine that resolves who depends on what, wrapped in an honesty architecture a buyer, a regulator or an auditor can trust. The next move is investability bite, one real parcel, one verified payor, one dollar, and we are raising to run it and to put this engine on the next corridors.
Built from live public data (USGS, EPA, USDA, GBIF) and real computation, with an artifact-QA audit that fails the deliverable if it asserts validated value for an unvalidated mechanism.
Built by Jay Gutierrez, PhD, working at the intersection of ecological network science, graph intelligence, nature-finance translation, and publishing the work that defines the structural-risk category.