Flying Rivers
The soy behind a UK portfolio depends on rain made two thousand kilometres upwind, by a forest no risk model prices. We follow that water back to its source, put a price on the dependency, and grade the source.
1 corridor · 324,542 ha · backtested across 22 harvest-years (IBGE, Brazil's official crop statistics) · every figure sourced
Half of Britain's nature risk starts abroad, and some of it blows in.
The Green Finance Institute puts up to 12 percent of UK GDP at risk from nature loss, and roughly half of it starts overseas. Part of it rides the wind, as moisture one forest releases and another region drinks in. Nature-risk platforms like GIST Impact and NatureAlpha price that exposure where a company's assets sit, mapping millions of locations against biodiversity and ecosystem data. That layer is real and useful. We find no tool that follows the rain the other way, upwind, to price the specific forest that makes it. The science is there, from precipitationsheds (the upwind land that supplies a region's rain) to Baker and colleagues' 2026 finding that Amazon rainfall is worth around twenty billion dollars a year, but no one has turned it into a number a risk desk can act on. That is the gap Fluvion fills, for the institutions holding South American soy exposure.
Three forests in Amazonas. 324,542 hectares.
Three standing forests in the western Amazon, 324,542 hectares between them. Every year they lift billions of tonnes of water into the air. The soy of southern South America, downwind, treats that as free rainfall: it has never paid for it, never insured it, and never asked whether the forest behind it is holding up.
About a fifth of the La Plata basin's rain comes from the Amazon.
We take the moisture flow from RECON, a published moisture-tracking dataset, and reconcile it to ERA5, the European reanalysis that reconstructs the global atmosphere hour by hour; running our own tracking model (WAM2layers) is the next milestone, not a claim we make today. The Amazon hands the La Plata basin 21.5 percent of its rainfall, inside the 12 to 35 percent that Zemp and colleagues published in 2014. This is a supply line into one of the world's great breadbaskets.
A pipe has one source. This has thousands, and their positions are not equal.
We have been drawing one lane, as if the rain had a single headwater you could fence off. Underneath that lane is the real object: the same reconciled moisture flows that give the corridor its number are a directed, weighted network across South America, thousands of cells that each drink rain and breathe some back, wired to the next. A source forest is not a patch of green in that picture. It is a position. Two forests that look identical from the road can sit in completely different places in the wiring, and the wiring, not the canopy, is what sets how much of the belt's rain leans on them. So the useful question stops being where does the water start, and becomes which positions carry the load.
Read it off the graph: a tenth of the upwind cells supply about half the belt's rain.
Point that network at one downwind economy, the South American soy grown in the Amazon-influenced rainfall system, and ask the inverse question the hub maps never ask: which upwind cells supply this belt's rain, and by how much of their own moisture. The dependency comes back concentrated, not spread. The top tenth of source cells supply about 48 percent of the belt's imported moisture, and it is not just the biggest evaporators doing the work, because the field barely tracks how much a cell evaporates. Reduce the whole field back to the one corridor we already validated and it returns the same weight the engine is built on, w = 0.2034, which is how we know the graph is not inventing anything.
Concentrated, yes. A hidden discovery, no.
We ran the field against its own null models before believing it, and it is worth showing what held and what did not. The concentration is real, and it is not an artifact of which cells evaporate the most. But most of it is the plain geometry of a corridor between a source and a sink: a shuffle that only preserves each cell's distance from the belt nearly reproduces the same concentration, and against equal-area patches drawn at random the field sits close to the middle of the pack. So read this as a sharper map of where the load already sits, not a new law of nature. The honest verdict is decision-useful, not Nature-grade, and we would rather say that plainly than let a clean map imply a precision it does not have.
The keystones hold. Only the load breathes.
The cells that carry the load do not change with the seasons, but how much they carry does. Three moisture models agree on the order of the keystone cells, while the total dependency swings about 3.3 times between the wet season and the dry, heaviest when the belt is wettest. Behind the single headline figure, w = 0.2034, the Amazon reaches the belt far more strongly in the wet season, while its share of the belt's rain runs the opposite way, near 27 percent in the dry season when little else falls. The dependence has a shape, and it is not a guess: the full field still reconciles to that committed annual value within 1.5 percent.
In 2022 the rivers ran thin, and the map turned red where it should.
The La Niña drought of 2021 and 2022 cut Rio Grande do Sul's soy by more than half while the cerrado (Brazil's central savanna) held steady. Our engine reproduces that split, which states fall and which hold, and repeats it on 2024, a year we held out a priori. The skill is real but modest: it recovers the sign and rank of which regions fall, not the exact loss, across a handful of drought-event-years. The numbers, held-out and pooled, are on the risk object below. This is a backtest, not a forecast.
As a rain machine, this forest is worth $350 a hectare.
That is the value of the harvests it quietly protects downstream, counted over thirty years. It already beats the $295 a hectare the land itself sells for, the regional average from Brazil's land registry. And it counts only water, only soy, only one place the rain lands. The full number is larger.
A forest losing resilience makes the same rain, less reliably.
The evidence points one way: a forest's condition does not change how much rain it makes on average, it changes how steadily it makes it. Intact, diverse forest is documented to hold its transpiration through drought, while degraded, resilience-losing forest wavers under the same heat. We read that condition from observed evaporation volatility and published resilience indicators, and we treat it as what it is: a descriptive, lagging correlate we are still validating, not a proven driver the price already carries. So it moves only the spread and the drought tail, never the $350, and it sits beside the price, never inside it.
A price today. A live signal next.
Today the engine prices one corridor: a moisture attribution we hold at medium-high confidence, and a drought backtest that reproduces the pattern. Next, we run our own WAM2layers moisture ensemble, fold in real trade-flow exposure to firm the UK number, and add a second corridor, each on the same published, auditable science. We already watch the source: the western Amazon has lost resilience for two decades, and the 2023 and 2024 droughts drove its evaporation sharply below normal. That sits beside the price, never inside it, until the link to the harvest is earned. A season-ahead warning is a research path, not a claim: Gate F is not run, and we publish the skill before we ever call anything a forecast. The direction is a graph, not a lane. The same load-bearing-node method extends from this one corridor toward a directed precipitationshed network, and it already runs on other systems: our Upper Colorado corridor applies the identical method to a river basin, and it is an honest sibling, because it reports where the load is not concentrated as readily as where it is.
The UK is the case study. The exposure is global.
Forest moisture underpins about 18 percent of the world's crop production, and croplands in 155 countries lean on forests in other countries for up to 40 percent of their rain (Wang-Erlandsson and colleagues, Nature Water, 2025). The South American soy that Britain imports feeds far more than Britain. We have priced one corridor for one case; the same upwind dependence runs under every economy below.
The United States is a competing exporter, exposed through the world soy price and its own domestic rainfall recycling, not through imports. How much of each economy's soy is specifically Amazon-fed is what Fluvion computes, corridor by corridor; the trade dependence here is context, from USDA and the European Commission.
Crop-moisture dependence: Wang-Erlandsson et al. 2025 (Nature Water). Rainfall value: Baker et al. 2026 (around twenty billion dollars a year for the Brazilian Amazon). Corridor benchmark: Zemp et al. 2014.
One corridor, validated. Two ways to back it.
One open core and one commercial layer, built by one team as a single venture. The open engine is the auditable standard underneath; the commercial layer prices it and serves it, to an investment desk or an AI agent, by API. The same engine reads two ways: it prices the dependency for the desks that carry the downside, and it grades the source for the capital that protects it. What we stand behind is the corridor's moisture link and its historical drought response; what we flag as scenario is the forward loss under deforestation and the source grade. Each corridor is a provenance-bound risk object like the one above, and the same integration and backtest extend it corridor by corridor: the first is done, the next are a repeatable campaign. The next move is yours.
- Validated. Reproduces the 2022 drought split, clears a held-out 2024 test at r = 0.45 (correlation), across 22 years of harvests.
- Sourced. Every number carries its source and confidence, from RECON moisture physics to peer-reviewed crop response.
- In the pack. The full provenance ledger, the backtest, and the live model, shared with serious counterparties on request.
jg@graphoflife.com · biome-translator.emergent.host
Built entirely from open public data and real computation. Every figure is indicative, not prudential grade. The full evidence pack, ledger and live model included, is shared on request. Phase 1B
Built by Jay Gutierrez, PhD, working at the intersection of ecological network science, graph intelligence, and nature-finance translation, and building the open engine behind Fluvion.