landscape-frontiers — geo-agent app

Interactive explorer for the biodiversity · climate · economy trade-offs of Polasky et al. 2026 (Science, "Landscape efficiency frontiers"). Draw an area or name a country → see where it sits in the three-objective space and how far it is from its efficiency frontier. Built on the geo-agent runtime (fork of geo-agent-template).

Files

• index.html — HTML shell, loads the geo-agent runtime + CSS from CDN (pinned @v3.17.1).
• layers-input.json — collections (carbon, ecoregions, WDPA, Overture admin), map view, draw_enabled, welcome examples. The objective surfaces (biodiversity composite, economic value, the optimal-land-use solution) are computed on demand via the duckdb-geo MCP and rendered with register_hex_tiles — not static toggles (the nci-frontiers economic layers are hex-parquet, no PMTiles).
• system-prompt.md — frontier framing + the canned SQL recipes (AOI scorecard, weight-sweep frontier, co-benefit finder, recolor-by-solution) + aggregation rules.

Status / dependencies

• Ingest complete (data-workflows #335, merged). nci-frontiers publishes economic revenue + grazing methane (native res 5); forestry, FLII, all 13 land-use transition costs, the six scenario-construction constraint masks, carbon-zones, and esa-lc-2025 (native res 8); plus the carbon-by-zone-lulc, predicts-crosswalk and predicts-sr-response lookups. Every layer carries an h5 parent for the paper's ~8000 ha decision unit. The whole CC0 Dryad input deposit is mirrored under raw/.
• Data readiness vs app fidelity — these differ, and the gap is the current work.
  • Catalog: economic and carbon objectives are gradable across all 13 alternatives (13 transition-cost surfaces; per-land-use carbon via carbon-zones × carbon-by-zone-lulc). The per-alternative biodiversity response is also gradable — predicts-crosswalk + predicts-sr-response are the paper's PREDICTS machinery, and the historical LULC epochs (band1/band14/band24) needed for the luage term are in the raw mirror. Five of the paper's six biodiversity sub-indices are constructible; only KBA is not (withheld from the deposit for licensing — data-workflows #334).
  • App: the canned frontier recipe in system-prompt.md is still a 2-alternative approximation (natural vs best-production) on all three axes. Upgrading it to the full 13-alternative argmax is a prompt change, not a data change — same separable SQL, more alternatives — and should go through a benchmark gate run before shipping.

Roadmap (see DESIGN.md)

• Charting (frontier curve + "you are here"): tracked upstream at geo-agent (opt-in generic chart primitive). Until then, frontier is delivered as a table + map recolor.
• Live weight slider (drag bio↔carbon↔economy → re-solve + recolor live): best as a generic geo-agent reactive-parameter control (relates to geo-agent #147 temporal slider); a frontier-specific extension can bolt on in the interim.

Research & provenance

The analysis behind this app (feasibility, sensitivity POCs, the efficiency-frontier reproduction) lives in research/. Data generation (H3 ingest of the CC0 Dryad layers) is owned by boettiger-lab/data-workflows (issues #334/#335).

Deploy

Two deployment targets, both supported from this repo:

Option A — GitHub Pages (bring-your-own LLM key)

Static hosting, no server. The llm block in layers-input.json (user_provided: true, OpenRouter default) means each visitor enters their own API key in the in-app settings panel — stored in the browser only, never on a server. The duckdb-geo MCP it queries is the public NRP endpoint.

• .github/workflows/gh-pages.yml deploys the repo to Pages on every push to main.
• One-time: Settings → Pages → Source → GitHub Actions. Live at https://boettiger-lab.github.io/landscape-frontiers/.

Option B — NRP / Kubernetes (lab LLM proxy, no user key)

k8s/ serves the static files via nginx and reverse-proxies /api/llm to the lab LLM proxy (the open-llm-proxy-secrets key lives only in the nginx config, never in the browser); config.json is injected at boot from k8s/configmap.yaml and overrides the BYO llm block. Live at https://landscape-frontiers.nrp-nautilus.io.

kubectl apply -f k8s/      # configmap + deployment + service + ingress (namespace: biodiversity)

Local preview

python -m http.server 8000   # open http://localhost:8000, enter an API key in settings