9  Evidence Base and Research Needs

Grassland birds are more abundant on planted solar sites than on the farmland around them, and that is known because people went out and counted them (Walston et al. 2025). A seeded ground layer under an array holding sediment and nitrogen out of the creek below is a claim of a different kind: it is the prairie-strip result from corn and soybean fields (Schulte et al. 2017), carried onto ground nobody has sampled. Both appear in this document, and the sentences that carry them look alike.

Every judgement this document makes about how well something is known lives here. The other chapters describe what agrisolar design is and what it can do; this one says what is established, what is inferred, and what nobody has measured. Keeping it in one place lets a reader weigh the whole document at once, and stops any chapter from quietly upgrading its own confidence.

It is arranged by remedy rather than by topic, because some gaps need a season of instrumentation, some need somebody to go to a primary table, and some need a project that does not exist yet to be built and reported.

9.1 What the evidence supports

Read this as the confidence behind each chapter, stated once rather than repeated as a caveat wherever a claim appears.

Water (§6.1). The dryland moisture link is verified (Barron-Gafford et al. 2019; Sturchio and Knapp 2025; Pinos et al. 2026), though bounded at some sites (Kannenberg et al. 2023), and infiltration and recharge at depth remain thin (Yavari et al. 2022). Interception is documented for prairie strips in cropland (Schulte et al. 2017), never yet from beneath an array; avoided loading is arithmetic rather than measurement, following from application rates on retired acreage rather than from a monitored stream.

Ecology (§6.2). Verified for grassland-bird gains (Walston et al. 2025), insect-community response (Walston et al. 2024), and bee diversity in pollinator-planted arrays (Bruninga-Socolar et al. 2025); scorecards codify the choices, though what each credits differs (Electric Power Research Institute 2021). The gain is context-dependent, largest where the surrounding landscape is already sparse in semi-natural habitat, so its size is set by what the site would otherwise have been rather than by the planting alone.

Agriculture and soils (§6.3). On the farming side, pollinator movement into adjacent crops is verified (Walston et al. 2024) and the farmland inside foraging distance has been mapped (Walston et al. 2018), while pest regulation is inferred, not measured at solar sites and a yield or input-cost change attributable to an array is not established anywhere. On the soil side, early gains are verified for ecovoltaic grassland restoration (Krasner et al. 2025), long-run accumulation is thinly measured (Carvalho et al. 2024), and the size of the construction debt is the least measured part of the ledger and among the likeliest to decide its sign, because the reported gains come from sites whose earthworks are not described.

Communities and economy (§6.4). The design principle is verified in the ecovoltaic frameworks themselves, where Tölgyesi et al. (2023) make a viable business model with embedded stakeholder engagement one of five co-equal pillars. The burden side is better measured than the benefit side (Bessette et al. 2024; Elmallah et al. 2023). Which structures deliver durable, fairly-shared benefit is still thinly evidenced.

The community scale, A3 (§5.3). Thin, and thinner than the confidence with which the middle is usually discussed. The capacity arithmetic is solid and federal. The delivery mechanism for pollination is measured, at two sites, for one function. Everything else is inference. Nobody has run the several-small-versus-one-large comparison at matched capacity and management, nobody has regressed land quality on project capacity at sub-parcel resolution, and no evaluation exists of whether the agricultural and pollinator adders in state incentive programs changed what actually got built. The outcomes footprint is under-evidenced at every zone on this transect, not only here. This section argues about where the decision still sits. It does not report where restoration has already happened.

The regional calibration (§2.3). The quantitative figures in this chapter are trace before citing: they come from secondary and program sources and have not been checked against primary records. The regional argument does not depend on any one figure.

What the siting studies show. Wu et al. (2026) is a study of utility-scale projects, at a 1 MW floor, and it treats each project as a single observation. It does not test whether siting drivers differ by project size, so it establishes what large-project siting responds to without establishing the contrast this section draws against smaller projects. That contrast still rests on Macknick et al. (2022) and Randle-Boggis et al. (2020) and on the archetypes, which is weaker evidence than a size-stratified analysis would be, and is why the peak of siting latitude remains an open question. Owusu-Obeng et al. (2025) is a preprint and has not been through review.

The prime-farmland argument. Studies that score solar facilities against soil quality find distribution-scale projects disproportionately on good land, and it is tempting to read that as careless siting. Two things complicate it.

The first is a confound. Community-scale solar goes where the load is, the load is where the people are, and the people are where the good soil was, because settlement followed arable land. Metropolitan areas hold a disproportionate share of the nation’s prime farmland for that reason, and development pressure has concentrated on the best soils for a century (Dillman and Cousins 1982). A fleet sited near load will sit on prime soil whether or not anyone was careless.

The second is resolution. A parcel classed as prime farmland is not uniformly productive. Across 338 US Midwest fields and 1,625 yield maps, about half of field area was high and stable, 32% was persistently low-yielding, and 18% swung with the weather, with the unstable ground concentrated in wet depressions (Maestrini and Basso 2018). Roughly a third of a prime field never performs. That 32% is exactly what siting latitude at A3 acts on, and a study that scores a facility by the soil class of its parcel cannot see whether the array took the good part or the bad part. The claim that mid-scale projects site carelessly and the claim that they site well are, at present, both untested at the resolution that would settle them.

9.2 Measurements nobody has made yet

Table 9.1: The measurements that would settle claims this document has to make on inference. Each is flagged where the chapters raise it, so a reader can check the claim rather than take the priority on faith.
Gap Flagged at What would close it
Nutrient and sediment interception under an array. Whether a seeded solar ground layer performs like a prairie strip. §6.1 Paired monitoring above and below an array sited on a flow path, against a strip or an unplanted control.
Recharge and the water balance at depth. Panel drip, altered cover, and what reaches the aquifer. §6.1 Measured drip and runoff, soil-moisture profiles, and coupled surface–groundwater modeling on instrumented sites.
Cumulative effect on one watershed. Whether arrays scattered through the same watershed add up, cancel out, or miss each other, and how much of the ground has to drain through them before a gauge can tell. §6.1 Gauged watersheds filling up with solar, against comparable ones without, and routing models built on those rather than one site at a time.
Soil carbon over decades. Early gains are measured; the multi-decade curve is not. §6.3 Re-sampling the early ecovoltaic sites on a decadal schedule, which needs a commitment longer than a grant cycle.
The construction debt, and how long it takes to repay. How much carbon mass grading costs a site, against how much a minimally graded build keeps. §6.3 Paired sampling before and after construction across projects with contrasting earthworks, which needs the earthworks recorded at the time.
Pest regulation from an array. Natural-enemy habitat is established for perennial refuge, never demonstrated from a solar site. §6.3 Natural-enemy and pest counts in fields at increasing distance from a planted array.
A yield or input-cost change attributable to the array. The step a farmer actually cares about. §6.3 Paired fields near and far from a planted array, with pollination and pest pressure measured alongside yield.
Whether a regional fleet of arrays functions as a connected network. Compatible seed mixes and calendars make one on paper; nothing establishes that dispersal actually happens between sites. §6.2 Mark–recapture or genetic work on a taxon whose dispersal distance is known, across arrays at a range of separations, with unbuilt patches as controls.
Whether co-located ground supports biodiversity the way unbuilt ground does. §6.2, and the stated gap in Brock et al. (2026) Comparative survey across co-located, conventional, and unbuilt sites in the same landscape.
Which benefit structures deliver durable, fairly-shared benefit. The instruments are known; their outcomes are not. §6.4 Longitudinal work on host communities rather than one-time acceptance surveys, with landowner and tenant reported separately.

The first of these is the one with the most resting on it. Interception is the water-quality route most projects could actually adopt, since it asks for a seed mix and a placement rather than for the retirement of a working field, and it is currently supported entirely by analogy.

9.3 Numbers that circulate without a traceable source

Some of this work is reading rather than measurement. The Sources chapter splits what has been checked from what has not, and the unverified block is not a list of doubts so much as a list of errands.

The grazing-cost figures are actively misquoted. Stewart et al. (2025) report McCall et al. (2023)’s $279/ha/yr as the median cost of sheep grazing, where in McCall’s Table 4 it is the median mowing cost at sheep-grazed sites; grazing itself ran about $50/acre/yr (§3.1). Both readings now coexist in the literature, and the per-hectare conversion compounds it. Anyone quoting a grazing cost should go to McCall’s table directly, and anyone who has already quoted the transposed figure should check which one they used.

The regional figures in §2.3 are the largest block of untraced material: irradiance ranges, queue and curtailment magnitudes, community-solar capacities, ground-coverage thresholds, and the acreages under SGMA and over the Ogallala. The argument in that chapter does not depend on any single one of them, which is why they were kept, but none should appear in a filing or a paper before it is traced. The community-benefit and grazing-economics material sits in gray literature for the same reason: real, widely practiced, and not yet synthesized anywhere citable.

9.4 Claims that a counterexample would test

The document makes several claims that could be wrong, and is more useful if the thing that would break them is named.

  • Sparing does not assemble at scale. §4.2 holds that S4 thins toward A5 because idle remnants are scattered. A utility-scale project assembled from scattered marginal remnants on ordinary farmland, rather than from uniformly pre-disturbed ground, would reopen that corner. None is documented here, and one would be worth reporting.
  • Both hand-offs fall at A3. Figure 5.1 places the structure-to-management crossing and the ownership-to-contracts crossing at the same zone, which is the load-bearing claim behind A3 being treated as the widest point for design. Whether the peak sits at A2 or A3 is argued from the archetypes rather than measured, and the curves are schematic. Moving the peak one zone left moves two claims, not one, so a set of real siting decisions that puts either crossing elsewhere would be worth reporting.
  • Regions resolve the pressures differently. The calibration in §2.3.2 is a set of hypotheses. A region doing something its profile says it should not is the most useful single contribution the chapter can receive.
  • The six levers are complete and non-overlapping. Every decision a project makes should land in exactly one of scale, siting, structure, management, ownership, and contracts. A decision that lands in none of them would mean the list is short; one that lands in all six would mean the categories are not cutting anything. The edge is the known partial exception, split deliberately between structure and management because a berm and a hedgerow fail differently (§1.2).
  • The A3 ceiling rests on one untested number. The zone tops out near 5 MW because a restorative community-scale array is taken to occupy at most about a third of its parcel (§5.3). That fraction is the load-bearing number in the whole scheme and has not been tested against built projects: at half a parcel the same fields carry 3 to 7 MW and the case for a higher ceiling partly returns. A set of built restorative community-scale arrays with their occupied fractions measured would settle it.
  • The vocabulary describes a site, and sparing is a claim about a landscape. Nothing here says where the array went instead. Statewide optimization shows the gap has teeth, since siting New York’s 2050 build to preserve farmland spares roughly 80% of the farmland a least-cost build would take and opens more than 41,000 hectares of forest doing it (Gallaher et al. 2026). The global models have the mirror-image problem: Brock et al. (2026) allocate land one sector to a pixel and say plainly that they could not represent co-location, which makes their land-demand figures upper bounds and leaves the space this lexicon describes invisible to them. A landscape-level accounting that can carry both would test whether the axis needs a companion term.
  • Positions are exclusive. The two-question reading rule (§4.1) should place any real project in exactly one position. A project that genuinely resists placement is a defect in the axis, not in the project.

9.5 What would make the evidence easier to build

These are recommendations about reporting, not about design. The document prescribes an aim and leaves the dimensions to the site, and none of what follows is a dimension.

  1. Report acreage alongside capacity. Megawatts are how a project is filed; acres are how the ground experiences it, and the two come apart with tracking and clearance (§3.1).
  2. Name the benefit delivery when claiming a function. A hydrologic claim that does not say whether it means moisture under the panels or nitrogen kept out of a creek cannot be checked, and the two need different evidence (§6.5).
  3. Separate the landowner from the tenant. Nearly 40% of Central Valley agricultural land is leased, and a conversion that sustains an owner can end the operation of the person farming the ground. Aggregate farm-income figures hide exactly the party most at risk (§6.4).
  4. Publish the counterfactual behind a sparing claim. Sparing is a claim about land that was not taken, so it is only checkable against the local market and the sites genuinely in play (§4.1).
  5. State the prior land cover and the earthworks when reporting soil carbon. Perennialization against a tilled baseline and conversion of existing grassland are opposite transactions, and the sign of the result depends on which one happened. How much the site was graded belongs alongside it, because a gain measured on ground that was never stripped and a gain measured on ground rebuilding from a stockpile are not the same number (§6.3).
  6. Give the region and the crop when reporting a shade-yield result. The sign flips with ambient temperature, so an unlocated yield figure travels further than the evidence behind it (§2.3.1).
  7. Report what did not work. Establishment failures, seed mixes that did not take, and grazing arrangements that fell apart are absent from the literature and present on the ground, and their absence is why cost expectations are unreliable (§3.1).