The Crop-Switching Problem

Every forest carbon project built on smallholder land carries an option it has not priced. The project sells long-dated removals. The household owns the land, and the land can be put to a different use on the strength of a price signal that arrives years before the first credit does. The standards call the consequence non-permanence, or reversal. What happens in the field is a crop switch, and it is normally well under way before anything shows up in a satellite image.

The framing that gets this wrong is the one that treats the baseline as the risk. Baselines model what the land would have done without the project: the counterfactual land use, its carbon trajectory, its opportunity cost. That is necessary work and, for afforestation and reforestation at least, it is now done reasonably well. It is also the wrong object. A baseline is a description of land. The risk is a description of a decision, taken repeatedly, by a household holding a portfolio and a discount rate, against a competing crop whose price series is in nobody's model.

The standards ask the question. They stop one step short.

Verra is more serious about this than its critics allow. The operative document is now the Procedure for Applying the AFOLU Non-Permanence Risk Tool, v5.0, issued 16 December 2025 under VCS Version 5; the opportunity-cost provisions discussed here carry over unchanged from v4.2 of October 2023. Internal risk carries an explicit Opportunity Cost category, scored on the net present value of the most profitable alternative land use against the NPV of the project activity — eight points where the alternative is expected to be at least 100 per cent more profitable, six above 50, four above 20, zero at or below 20, with negative scores where the project itself is materially more profitable. The tool goes further than most readers notice and requires that the NPV analysis also compare potential price fluctuations of commodities impacted by the project.

That requirement is real, and it is under-specified. No distribution is named. No volatility parameter, no elasticity, no minimum number of scenarios, no guidance on which price series. In practice it resolves into a single deterministic ratio, sorted into four steps — twenty, thirty and fifty percentage points wide, and then one open-ended step covering everything above a doubling — computed once at validation and revisited only at a verification event. The question is asked. The answer is a point estimate on a moving target.

Two structural gaps matter more than the parameterisation.

• The subsistence route. The first question in the opportunity cost table asks whether baseline activities are subsistence-driven. Where most of them are, no NPV analysis is required at all; the project instead demonstrates net positive community impacts, scoring zero where demonstrated and eight where not. The classification is binary, and the tool puts the onus on the proponent to substantiate what counts as a credible alternative. The common case in a smallholder mosaic is neither pole: a household growing food for itself on part of the holding and one or two highly price-elastic cash crops on the rest. Classified as subsistence, the most price-exposed projects in the portfolio are routed around the price test entirely.
• Exit is already priced as a total loss, but not as a probability. Grouped projects whose contracts with individual participants run shorter than the project longevity must apply the weaker of the two longevity formulas, and where permanence monitoring of a departed participant is not possible, the tool requires the project to assume a complete loss of carbon stocks when that instance leaves. So the consequence of a crop switch is already written into the rules at full severity. What is missing is the likelihood, and the likelihood is a function of a commodity price.

Set against that, the commitment horizon is long. A project requesting registration on or after 1 January 2024 fails the risk assessment outright if project longevity is under 40 years, and the tool states that anything below 40 years is ineligible for the Core Carbon Principles label. Forty years of commitment, and participant agreements that are commonly a fraction of it.

A twenty-five-fold premium that became a two-fold premium

Uganda's vanilla sector is the cleanest recent illustration, because both sides of the ratio moved.

Uganda's Daily Monitor reported in June 2026 that through 2010 to 2016 the crop hit an average of UGX 350,000 per kilogramme — about twenty-five times the price of a kilo of arabica coffee — and now commands about UGX 20,000, around twice the price of coffee. Nearly 70,000 smallholders across nearly 40 districts are in the crop. Nothing about the plant changed. The denominator changed and the numerator collapsed.

The collapse is imported. Madagascar exported about 4,300 tonnes in 2023/24 against global demand of only about 2,500 tonnes, and average export prices fell 40.2 per cent to USD 38 per kilogramme in the fourth quarter of 2025, against USD 63.6 in the fourth quarter of 2024, according to Ecofin Agency. A Ugandan grower is a price-taker in a market cleared in Madagascar.

Coffee went the other way. Arabica futures topped USD 4.00 per pound for the first time in the session of 5 February 2025, adding to gains of 70 per cent across 2024, as reported at the time, and peaked at USD 4.41 in early February. Uganda shipped 8.8 million bags worth USD 2.5 billion in the twelve months to February 2026, volumes up 41 per cent and value up 61 per cent year on year, with February 2026 farmgate arabica parchment at UGX 15,500 per kilogramme and robusta FAQ at UGX 12,250, per Nile Post.

A vanilla agroforestry project written in 2014 would have scored close to zero on opportunity cost, and correctly. The same project assessed in 2026 faces a competing crop that is plausibly the more profitable alternative land use. No tree died. No tenure changed. A price series in another country changed, and a second price series moved the other way at the same time.

The feedback loops are worth naming, because they are what turns a price move into a permanence event. In the boom, theft was severe enough that growers harvested immature beans ahead of thieves, which cost quality, which cost price. The state now sets the harvest window by directive: the State Minister for Agriculture declared the 2023 season B harvest open from 15 January to 15 April 2024, reported here. A government legislating harvest dates is managing a price-behaviour problem, not an agronomy problem, and that is a signal a due-diligence visit will not produce.

The asymmetry that is not in the model

The economics of this are well established and almost never imported into carbon project design. Siegle, Astill, Plakias and Tregeagle's 2024 review in Agricultural Economics, Estimating perennial crop supply response, sets out the structure: because the biological growth process of trees makes the stock of orchard capital slow to adjust, perennials exhibit asset fixity, so short-run response (within-year adjustments) and long-run response (plantings and removals, which change bearing acreage) are different animals. Planting and removal are separate decisions on separate lags. And because the investment is largely irreversible, adjustment does not happen where expected net revenue equals opportunity cost, but at a threshold high enough to compensate for the risk — the real-options framing applied to perennials by Price and Wetzstein in 1999, whose model of peach trees as an irreversible sunk-cost investment shows, in the review's words, large potential gaps between the expected profits that would spur entry and exit.

That cuts both ways. An established stand sits inside a hysteresis band, and a modest adverse move in the price ratio will not clear the exit threshold — which is protective. But entry faces the same band, and that is the half developers underestimate: an offer that merely matches the NPV of the competing crop does not get planted. It has to clear a threshold set by the volatility of the alternative, not by its mean.

The deeper problem is that the observable sequence runs backwards from the one being monitored.

Horizon	What the household does	What the project can see
Weeks	Reallocates labour and cash inputs toward the crop with the better ratio	Nothing. Land cover unchanged
One season	Defers pest and disease treatment, prunes less, harvests differently	Yield and quality drift; pest severity rising
One to three years	Interplants or underplants the competing crop	Canopy signal ambiguous or improving
Three to five years	Replacement completes; establishment lag of the new crop runs out	Land-cover change finally legible
At contract renewal	Does not renew, or leaves the group	Complete loss of stocks assumed where the instance cannot be monitored

Remote sensing catches row four. The decision was made in row one.

Maintenance is the leading indicator

The best evidence for that sequence comes from a disease literature rather than a carbon one. Avelino and colleagues, writing in Food Security in 2015 on the coffee rust crises in Colombia and Central America, 2008–2013, found that all the intense epidemics experienced in the region during the preceding 37 years were concurrent with periods of low coffee profitability — driven by price falls, as in Central America in 2012–13, or by increases in input costs, as in Colombia from 2008 to 2011. Their conclusion is blunt: "Low profitability led to suboptimal coffee management," and the weakened plants were more vulnerable. Central American production fell 16 per cent in the 2012–13 harvest against 2011–12 and a further 10 per cent the following year; Colombian production fell 31 per cent on average across 2008 to 2011 against 2007.

The early-warning system they propose uses the profitability of the coffee crop as its economic indicator, on the stated reasoning that profitability is a proxy for the investments in inputs and labour that farmers will make during the year. Read in reverse, that is a permanence monitoring instrument. The chain runs price, then management, then biology, then yield, then land use. Disinvestment is visible in pest and disease behaviour one to two seasons before it is visible in canopy.

Which means the cheapest permanence instrument available is not imagery. It is asking growers what they are still treating, what they have stopped treating, and who told them to.

That is the design logic behind two datasets we hold. The first is farm-level cost, yield and income data across the Trifinio region of Guatemala, Honduras and El Salvador for coffee years 2008–09 through 2012–13 — five consecutive years running directly into the rust epidemic. Its value is structural rather than headline: it holds cost alongside yield in the same farm record, across exactly the profitability trough the literature identifies, which is what you need to see input withdrawal and yield response as one movement rather than two.

The second is a 2013–14 grower and trader survey across Buikwe, Mukono and Kamuli districts in Uganda. It was built around four instruments: GPS-located farms, so plot identity survives a decade; ranked pest severity, which is the maintenance signal above; elicited risk preference; and named advice networks, because a grower's decision to switch is rarely taken alone and the network is the transmission mechanism. We report the instrument design here, not per-farm results. The point is that all four are cheap to collect, repeatable, and move before land use does.

Loss aversion, not risk aversion

One correction to how this is usually discussed. The variable in most project narratives is risk aversion, and it appears to be the wrong parameter.

Kijima's 2019 study in PLOS ONE combined panel data from 539 Ugandan farming households across five districts, surveyed in 2010 and 2017, with risk and loss preferences elicited through lab-in-the-field experiments. Loss aversion predicted the outcome: more loss-averse households were less likely to grow rice and less likely to expand the area under it, and allocated more land to root crops such as cassava and sweet potato. Risk aversion showed no evidence of influencing the decision to grow rice at all.

Joining a carbon-linked agroforestry programme is an adoption decision, so the binding parameter is loss relative to a reference point, not variance around a mean. A contract that removes a household's ability to capture the next price spike is experienced as a loss, even where its expected value is higher — and the reference point is set by the last boom, which in vanilla-growing Uganda is living memory. Three design consequences fall out of that directly: leave the upside participable rather than fixed; do not ask for exclusivity across the whole holding when part of it will do; and make the near-term payment the certain one, with the long-dated payment carrying the upside share. Most carbon contracts are built the other way round.

The operator half

The other half of this is what happens when you stop advising and start selling.

Our vanilla agroforestry pilot in Uganda became a trading business. It sits on a nineteen-year working relationship with one community organisation, the Rural Agency for Sustainable Development in Nkokonjeru, Mukono district, whose director sits on our advisory board. That relationship is the reason we have a view on this at all, and it is also the reason we are cautious about generalising from it: one crop, one district, one trading business.

A trading business in that crop is a price-taker on a benchmark cleared in Madagascar. That is the grower's position exactly, and it is structural rather than cyclical: competing on nothing but price, with no instrument for changing the price. The only move available is to sit further along the chain than the raw material, where the product carries its own demand and the margin is not set by export prices in another country.

Read that back into permanence and it is not a marketing observation. If all a project offers a household is a buyer for raw material at the prevailing price, the project has no answer at all to a competing crop paying more this season — it is one bid among several, and a weak one. If the project sits inside a chain that captures processed-goods margin and shares it back, the household's reference point moves: the agroforestry crop stops being a commodity bet and becomes a supply relationship with a floor under it. That is a permanence intervention. It does not appear anywhere in a risk tool, and it is doing more work than the buffer contribution.

What to model

A competing-crop screen is not expensive. It is mostly reordering work that a feasibility process already does badly.

Test	What it requires	Where it usually fails
Name the competing crop	Not the baseline land cover — the crop a household in this district would plant if its price doubled	Answered from national statistics rather than traders and extension officers
Relative price series	Ratio of project crop to competing crop, farmgate, local currency, ten years minimum	Only the project crop's price is collected, and in USD
Historic switch threshold	The ratio at which this district last switched. There is almost always a last time	Nobody asks, because it predates the project's own records
Lags in both directions	Establishment period of the competing crop; removal cost and time for the project crop	Only the project crop's establishment lag is modelled
Maintenance elasticity	What input and labour spend does at each ratio — cost data at farm level, not yield alone	Monitoring collects yield and canopy, which lag the decision
Contract horizon against project longevity	The gap between participant agreement length and the 40-year longevity requirement	Treated as an administrative detail, not an exposure
Exit consequence	The complete-loss assumption for unmonitorable departures in grouped projects	Priced at severity, never at probability

Two notes on what already helps, and how far. VM0047 v1.1, active since 14 May 2025 (v1.0 from 28 September 2023), sets a dynamic performance benchmark under its area-based approach from remotely sensed matched control plots, and establishes the crediting baseline at every verification. That genuinely handles part of this: a region-wide price shock that moves everyone's land use moves the control plots too, and the benchmark absorbs it. What it cannot absorb is the idiosyncratic case, where project participants exit and controls do not. Dynamic baselines are good insurance against correlated crop-switching and no insurance at all against individual withdrawal.

And the buffer is calibrated to the project's own risk score, not to the volatility of the crop it is competing with. Two projects with identical agronomy, tenure and management, one next to a cocoa frontier and one next to a cassava field, are not carrying the same permanence risk. The tool has no way of saying so.

The claim

Model the competing crop, not just the baseline. The baseline tells you what the land was doing. The competing crop tells you what the household will do when something better comes along — and something better arrives on a commodity cycle, not on a crediting period.

This is not an argument that smallholder agroforestry carbon is unbankable. It is an argument that the risk is legible, cheap to screen for, and currently assessed once, deterministically, on a scale whose top step is open above a doubling. A project that can name its competing crop, produce the relative price series, and state the ratio at which the last switch happened in that district is a materially better underwriting proposition than one that cannot. Very few can.