How Refuge builds your Native Blueprint

Your address is geocoded to a single point, then resolved against three datasets in parallel:

• County, via PostGIS point-in-polygon against TIGER/Line 2024 boundaries.
• EPA Level IV ecoregion — one of 180 fine-grained ecological units that subdivide California by climate, geology, soils, and historical vegetation. A Level IV ecoregion is finer than a USDA hardiness zone: two yards 30 miles apart can sit in entirely different ecoregions, and the native plant communities of each are different.
• Property context — elevation (USGS), soil series and drainage (USDA SSURGO), 30-year climate normals (PRISM 4km), surface-water proximity (USGS NHD), watershed (USGS Watershed Boundary Dataset HUC12, walked up to sub-region and terminal water body), USDA hardiness zone (PRISM 2023), and average last spring frost (NOAA NCEI 1991-2020 normals, nearest station).

Before any plant is considered for your report, it must pass two native-range gates and two practical filters.

• State-level gate (NatureServe). Where NatureServe Explorer has assessed the species, it must list California among its native states; species it hasn’t assessed aren’t blocked here — the stricter county gate below is the real test. This catches obvious wrong-state errors.
• County-level gate (AI-curated, multi-vendor verified — Step 3 below). The species must be documented as genuinely native to your specific county, not just somewhere in California.
• Tier: keystone or supporting species only. Tier 3 ornamentals with limited wildlife contribution are excluded.
• Commercial availability: only plants stocked by California native-plant nurseries, reconciled against Calscape’s current carrying-nursery data. Recommending a plant a customer can’t source doesn’t help.

The county-level gate is the V1 architectural advance and the part of the pipeline that deserves the most explanation. A species being “California-native” is not enough — California spans coastal redwood forest, alpine meadow, oak savanna, chaparral, and four deserts. Recommending a Sierra Nevada montane species for a downtown Sacramento yard is the kind of error that makes a report feel auto-generated. We built a four-pass pipeline specifically to eliminate it.

1. Per-species pass. A state-of-the-art language model lists every California county where each species in the candidate pool is genuinely native, drawing on training data covering Jepson eFlora, CNPS Calscape, BONAP, and USDA PLANTS. Each cell is tagged with a confidence level (core / edge) and a brief habitat note.
2. Precision veto. A second state-of-the-art language model from a different vendor reviews every listed (species, county) cell with a directed “is this actually native?” prompt. Cells the second model confidently rejects are dropped. Because the two models are built independently, this cross-check catches errors a single model would repeat — a cell that survives the veto has agreement from two independent passes rather than one.
3. Reverse-direction sweep. The first model is queried again, this time county-first: “of all the species in the recommendable pool, which are genuinely native to this county?” This catches species the per-species pass missed — different framing, different errors.
4. Adjudication. A third model resolves disagreements between the per-species and reverse-direction passes, plus reinstates cells the precision veto dropped that the third model can confirm are real natives. Only adds that the adjudicator confirms with high confidence are accepted.

The resulting dataset has explicit per-cell provenance: which model said what, with what confidence. No hidden manual lists; every choice is traceable. Audited against well-known California native plants across launch counties; verified to exclude known wrong-region contamination (redwoods don’t appear in Sacramento Valley reports; giant sequoias don’t appear in coastal addresses; coast live oak doesn’t appear in inland desert reports).

The candidate pool from Steps 1–3 is then filtered against your specific address:

• Elevation — species whose elevation range doesn’t overlap your property are excluded, with a 300-foot buffer in each direction. Species that the per-species curator vouched for your specific county get an extended 500-foot buffer — letting foothill-fringe species appear for sea-level addresses near their range edge.
• Substrate exclusion — species whose plant communities don’t fit your soil are excluded. Serpentine endemics are excluded on non-serpentine soil; wetland obligates are excluded on upland soil; coastal-strand species are excluded inland.
• Synonym suppression — known taxonomic synonyms are collapsed to their canonical species so a report doesn’t show two cards for the same plant under different botanical names.
• Climate match — most species carry a climate niche: the precipitation and temperature range of their California occurrence records, from PRISM 4km normals. A species whose niche doesn’t overlap your property’s 30-year climate is excluded. This is the desert / montane discriminator — it’s why a Palm Springs report leads with mesquite and palo verde, not coastal oaks.
• Local corroboration — a species with no climate niche on file must instead be backed by verified GBIF occurrence in your ecoregion. A county-native still surfaces where observation density is thin, but a name with neither a matching climate niche nor local records is held back rather than shown on faith.

Within the eligible pool, species are ranked by a composite wildlife score that combines four taxonomically distinct signals, so a plant’s rank reflects its impact on insects, native bees, broader pollinators, and birds — not just one of those.

The four signals and their weights:

• Caterpillars (weight 0.50) — California-specific counts of moth and butterfly species whose larvae feed on this plant, sourced from CNPS Calscape where available, falling back to Doug Tallamy’s national genus-level data for the ~15% of species without a Calscape entry. The keystone-genus framing (oaks, willows, cherries, asters, goldenrods support an outsized share of the insect biomass that feeds nesting songbirds) anchors the whole approach.
• Specialist native bees (weight 0.25) — per-genus count of California-distributed oligolectic bees (bees that depend on a single plant genus or family for pollen) from Jarrod Fowler’s 2020 compilation. Highest for late-season composites — goldenrods, gumplant, asters — and zero for wind-pollinated genera like oaks and pines.
• Pollinator breadth (weight 0.15) — per-species count of distinct flower visitors (bees, hoverflies, beetles, butterflies, and more) from GloBI (the Global Biotic Interactions network). Captures pollinator diversity beyond specialists, log-normalized to blunt the sampling bias toward common showy plants.
• Bird food (weight 0.10) — a hand-curated tag (berries, acorns, nuts, seeds, samaras) for the California natives known to provision birds directly. There is no defensible public dataset for species-level CA bird-plant relationships, so we curate rather than fake it.

Two bonus modifiers:

• Keystone bonus (+0.10) for Tier 1 (keystone) species, so the highest-impact plants get a soft preference inside their layer.
• Confidence bonus (+0.05) for species the per-species curator marked as “core” in your county (vs. “edge”), so the most clearly-native plants surface slightly higher.

Each plant card surfaces the underlying signals as wildlife icons — a caterpillar, bee, bird, and hummingbird mark — so you can see at a glance why a plant ranks where it does. The exact caterpillar-host and specialist-bee counts live on each plant’s detail view. The composite score sets the default order; the icons, with that full detail, are the audit trail.

Species are grouped into five structural layers so the list reads as a complete planting palette, not a wall of one growth form: trees, shrubs, perennials & wildflowers, ground covers, and native grasses.

Within each layer, a genus cap keeps the list diverse: three species per genus for keystone genera (oaks, willows, manzanitas, ceanothus, pines, currants, salvias, penstemons, eriogonums, lupines, monkeyflowers — the regional anchors most customers expect to see well-represented), two per genus elsewhere. Inside a genus the cap picks the species that’s actually most abundant in your ecoregion, not the one with the highest state-level wildlife score — so a Sacramento Valley address gets Valley Oak, not the foothill-dwelling Blue Oak with a near-identical score but a fraction of the local presence.

Final ordering within the report: layer → composite wildlife score → commercial availability (widely-stocked plants edge out specialty-only ones at the same wildlife rank) → botanical name.

The plant list is the core of the report, but Refuge also renders four pieces of context that frame why these species and not others:

• Aerial view of your land — a satellite tile centered on your geocoded coordinates so the report begins with the actual place, not an abstraction.
• Property data panel — elevation, soil series, USDA hardiness zone, 30-year rainfall and temperature, and average last spring frost. Each field hides cleanly when its source data is unavailable.
• Watershed cascade — your USGS HUC12 sub-basin, walked up to the sub-region and terminal water body (Pacific, San Francisco Bay, Sacramento Delta, Salton Sea, etc.). Native plantings’ deep roots help rain soak in rather than run off the way it does over shallow-rooted lawn, keeping pollutants out of the rivers below.
• Property narrative — one short paragraph written for your specific property by a large language model under a strict accuracy contract: it builds only from your ecoregion’s vetted facts — the characteristic wildlife, plant communities, and keystone plant of your land — naming only wildlife from that vetted list and using only the provided host-plant numbers. It doesn’t recite your elevation or soil figures (the data panel shows those) and doesn’t pull from web search; it may add one widely-known natural-history detail about a featured species. It places your land in its ecoregional context in the same authoritative-naturalist voice as the hand-written ecoregion narrative below.
• Nearest native nurseries — the closest three California native-plant retailers within 50 miles, drawn from a curated directory of CNPS-listed nurseries.

Every Refuge recommendation traces back to public data. The full source list:

• EPA Level IV Ecoregions of California — United States Environmental Protection Agency via USGS.
• TIGER/Line 2024 County Boundaries — U.S. Census Bureau; used for the point-in-polygon county lookup.
• NatureServe Explorer — state-level native-presence records; the Layer-0 native-range gate.
• County-level native-range curation — AI-curated dataset using multi-model verification, built on training-data knowledge of Jepson eFlora, CNPS Calscape, BONAP, and USDA PLANTS.
• GBIF occurrence records — Global Biodiversity Information Facility, research-grade observations.
• CNPS Calscape — per-plant pages; primary California-specific lepidoptera host count for most of the recommendable pool (the rest fall back to the national Tallamy data below).
• Tallamy & Shropshire 2009/2020 — national genus-level lepidoptera host data via the National Wildlife Federation Native Plant Finder; the lep fallback for species without a California-specific count.
• Jarrod Fowler 2020 — Pollen Specialist Bees of the Western United States; the specialist-bee signal.
• GloBI (Global Biotic Interactions) — flowersVisitedBy records; the pollinator-breadth signal. Open data under CC0.
• Bird-food categorization — hand-curated list of ~80 California natives with documented direct bird-foraging value, compiled from CNPS, Audubon, and ornithology references.
• Calflora and Jepson eFlora — native status, growth form, horticultural traits.
• USDA NRCS SSURGO — soil series and drainage.
• USGS Elevation Point Query Service — per-address elevation.
• PRISM Climate Group — 30-year climate normals (4km grid).
• USGS National Hydrography Dataset — surface-water proximity.
• USGS Watershed Boundary Dataset (HUC12) — sub-basin lookup and the parent-region cascade drawn at the top of every report.
• PRISM Climate Group + USDA — 2023 USDA Plant Hardiness Zone Map (800m grid).
• NOAA NCEI 1991-2020 Climate Normals — per-station median spring and fall freeze dates, growing-season length.
• CNPS native plant nursery directory — the curated list of California native-plant retailers feeding the "Where to buy" section.
• Mapbox Satellite — aerial imagery for the report’s opening view, attribution displayed beneath the image.
• iNaturalist and Wikimedia Commons — research-grade photographs and per-photo attribution, used under their Creative Commons licenses with photographer credit on each image.

iNaturalist and Wikimedia Commons images are used under their Creative Commons licenses (CC0, CC-BY, CC-BY-SA, public domain), displayed unmodified with photographer credit. Embedding within Refuge reports is not a derivative work.

Refuge is a vetted shortlist, not a landscape design. Placement, hardscape, shade, succession, and cultivar selection are all out of scope — where the wild form differs substantially from common trade cultivars, the species card notes it. The list starts you right — your eyes on the ground finish it.

Refuge is not a substitute for professional landscape design, arboricultural consultation, or fire-defensible-space planning.

V1 is calibrated for Mediterranean and Valley California — the climate zones most California gardens occupy. Desert and high-Sierra addresses are supported too: a PRISM climate filter keeps each report climate-appropriate (a Palm Springs address leads with mesquite, palo verde, and ironwood — not coastal or montane trees), and hyperarid salt-flat sites with too few cultivable natives return an explicit limited-coverage notice rather than a forced list. These climate types saw less validation than the Mediterranean core, so we’re continuing to refine them post-launch with signal from real customer reports.