woozyking · Aug 25, 2026 · Aug 24, 2026
diff --git a/.agents/skills/crc-assess-agricultural-climate-risk/SKILL.md b/.agents/skills/crc-assess-agricultural-climate-risk/SKILL.md
@@ -0,0 +1,47 @@
+---
+name: crc-assess-agricultural-climate-risk
+description: Assess bounded USDA crop-cover or global FTW field evidence against CRC canonical flood or drought hazards. Use for agricultural sourcing, crop exposure, field screening, and open climate-risk baselines; not for yield prediction, ownership, or a synthetic farm-risk score.
+---
+
+# CRC agricultural climate risk
+
+Produce an auditable agricultural evidence layer and a separate per-hazard
+evaluation. Never infer ownership, yield, production, or a combined risk score.
+
+## Intake and source choice
+
+Reuse a supplied geography, year, crops, hazard, and decision context. If the
+target or decision is absent, ask: “Which sourcing area or agricultural region
+should I assess, and what decision should the screen inform?” Resolve and show
+a narrow WGS84 AOI.
+
+- In CONUS, default to USDA CDL 30 m and the latest complete requested year.
+  Use 10 m only for 2024–2025 when the added detail matters.
+- Elsewhere, use FTW field boundaries for the requested ISO country and retain
+  continuous confidence. State that crop type is unavailable.
+- Default flood to EFAS in Europe and GloFAS elsewhere. Add EDO SMI only when
+  drought is relevant and covered.
+
+## Workflow
+
+1. Read [the agricultural playbook](../../../ai-playbooks/agricultural-climate-risk.md).
+2. Materialize each applicable hazard into its own canonical Parquet and read
+   its embedded source, unit, semantics, pathway, horizon, tail, support, and
+   H3 resolution.
+3. Run `pipelines/agricultural_climate_pipeline.py` with a bounded AOI. Let the
+   hazard metadata choose H3 resolution. The notebook twin is
+   `notebooks/agricultural_climate_risk.ipynb`.
+4. Preserve USDA year/class and sampled crop share, or FTW year/confidence/area.
+   Preserve match status and missingness in each hazard evaluation.
+5. Keep hazards with incompatible units, horizons, or semantics separate.
+   Flag return-period extrapolation and low-confidence or missing agricultural
+   coverage.
+
+## Deliverable
+
+Return the agricultural-units Parquet, one canonical evaluation Parquet per
+hazard, map and summary Parquets, a PMTiles presentation artifact, and a
+manifest with versions, filters, hashes, row counts, metadata, and warnings.
+Show a map plus a per-crop or confidence-stratified exposure summary. Label
+USDA crop share as sampled pixels—not acreage—and FTW polygons as
+remote-sensing field units—not parcels or proof of ownership.
diff --git a/.agents/skills/crc-assess-agricultural-climate-risk/agents/openai.yaml b/.agents/skills/crc-assess-agricultural-climate-risk/agents/openai.yaml
@@ -0,0 +1,5 @@
+interface:
+  display_name: "CRC Agricultural Climate Risk"
+  short_description: "Screen crop and field layers against hazards"
+  default_prompt: "Use $crc-assess-agricultural-climate-risk for my sourcing region."
+
diff --git a/AGENTS.md b/AGENTS.md
@@ -14,6 +14,7 @@ repository discovery; see README.md's "Platform discovery" section.
 | crc-screen-mortgage-flood | [.agents/skills/crc-screen-mortgage-flood/SKILL.md](.agents/skills/crc-screen-mortgage-flood/SKILL.md) |
 | crc-model-flood-insurance-loss | [.agents/skills/crc-model-flood-insurance-loss/SKILL.md](.agents/skills/crc-model-flood-insurance-loss/SKILL.md) |
 | crc-assess-asset-portfolio-risk | [.agents/skills/crc-assess-asset-portfolio-risk/SKILL.md](.agents/skills/crc-assess-asset-portfolio-risk/SKILL.md) |
+| crc-assess-agricultural-climate-risk | [.agents/skills/crc-assess-agricultural-climate-risk/SKILL.md](.agents/skills/crc-assess-agricultural-climate-risk/SKILL.md) |
 | velo-underwrite-property-climate | [.agents/skills/velo-underwrite-property-climate/SKILL.md](.agents/skills/velo-underwrite-property-climate/SKILL.md) |
 | velo-triage-portfolio-insurability | [.agents/skills/velo-triage-portfolio-insurability/SKILL.md](.agents/skills/velo-triage-portfolio-insurability/SKILL.md) |
 | velo-assess-company-climate-risk | [.agents/skills/velo-assess-company-climate-risk/SKILL.md](.agents/skills/velo-assess-company-climate-risk/SKILL.md) |

diff --git a/README.md b/README.md
@@ -42,6 +42,7 @@ Note: to run an enterprise twin, connect the [`CDT Express MCP server`](https://
 | **Mortgage / collateral** | **Start:** `Use $crc-screen-mortgage-flood.`<br>**Optional:** `Use $crc-screen-mortgage-flood for Toronto, Canada.` | **Start:** `Use $velo-underwrite-property-climate.`<br>**Optional:** `Use $velo-underwrite-property-climate for 392 Markham Street, Toronto.` |
 | **Insurance** | **Start:** `Use $crc-model-flood-insurance-loss.`<br>**Optional:** `Use $crc-model-flood-insurance-loss for Rotterdam using my attached portfolio and approved depth-damage curve.` | **Start:** `Use $velo-triage-portfolio-insurability.`<br>**Optional:** `Use $velo-triage-portfolio-insurability for Example Insurance Holdings.` |
 | **Corporate finance / investment** | **Start:** `Use $crc-assess-asset-portfolio-risk.`<br>**Optional:** `Use $crc-assess-asset-portfolio-risk for Frankfurt, with flood and drought.` | **Start:** `Use $velo-assess-company-climate-risk.`<br>**Optional:** `Use $velo-assess-company-climate-risk for the S&P 500.` |
+| **Agricultural sourcing / crop exposure** | **Start:** `Use $crc-assess-agricultural-climate-risk.`<br>**Optional:** `Use $crc-assess-agricultural-climate-risk for corn and soybeans near Ames, Iowa.` | Use the company/portfolio twin when ownership, suppliers, values, or enterprise scenarios are required. |
 
 After running both sides for the same real target:
 
@@ -90,6 +91,7 @@ both outputs exist.
 | Mortgage / collateral | [`crc-screen-mortgage-flood`](.agents/skills/crc-screen-mortgage-flood/SKILL.md) → [`velo-underwrite-property-climate`](.agents/skills/velo-underwrite-property-climate/SKILL.md) |
 | Insurance | [`crc-model-flood-insurance-loss`](.agents/skills/crc-model-flood-insurance-loss/SKILL.md) → [`velo-triage-portfolio-insurability`](.agents/skills/velo-triage-portfolio-insurability/SKILL.md) |
 | Corporate finance / investment | [`crc-assess-asset-portfolio-risk`](.agents/skills/crc-assess-asset-portfolio-risk/SKILL.md) → [`velo-assess-company-climate-risk`](.agents/skills/velo-assess-company-climate-risk/SKILL.md) |
+| Agricultural sourcing / crop exposure | [`crc-assess-agricultural-climate-risk`](.agents/skills/crc-assess-agricultural-climate-risk/SKILL.md) plus the [agricultural playbook](ai-playbooks/agricultural-climate-risk.md) |
 | Reconciliation / sales engineering | [`compare-crc-velo-assessments`](.agents/skills/compare-crc-velo-assessments/SKILL.md) |
 
 For an open baseline without supplied files, the AI should:
@@ -202,6 +204,7 @@ These workflows acquire and canonicalize hazard data for a selected geography.
 | [`notebooks/flood_risk_by_province.ipynb`](notebooks/flood_risk_by_province.ipynb) | [`pipelines/flood_admin_pipeline.py`](pipelines/flood_admin_pipeline.py) | EFAS area selection → pinned AOI cache → canonical curves → H3/admin join → Overture places → PMTiles |
 | [`notebooks/jrc_global_flood_hazard.ipynb`](notebooks/jrc_global_flood_hazard.ipynb) | [`pipelines/jrc_flood_pipeline.py`](pipelines/jrc_flood_pipeline.py) | EFAS/GloFAS selection → pinned AOI cache → fitted curves → return-period evaluation |
 | [`notebooks/jrc_edo_drought_index.ipynb`](notebooks/jrc_edo_drought_index.ipynb) | [`pipelines/jrc_drought_pipeline.py`](pipelines/jrc_drought_pipeline.py) | EDO SMI area/year selection → annual-minimum AOI cache → lower-tail curves → support-aware evaluation |
+| [`notebooks/agricultural_climate_risk.ipynb`](notebooks/agricultural_climate_risk.ipynb) | [`pipelines/agricultural_climate_pipeline.py`](pipelines/agricultural_climate_pipeline.py) | Bounded USDA crop samples or FTW field polygons → canonical hazard evaluation → source-aware maps and weighted summaries → PMTiles |
 
 ### Asset portfolio analytics
 

diff --git a/ai-playbooks/agricultural-climate-risk.md b/ai-playbooks/agricultural-climate-risk.md
@@ -0,0 +1,72 @@
+# Agricultural climate-risk playbook
+
+Use this playbook to answer location-specific questions such as:
+
+- Which sampled crop classes in an Iowa procurement area overlap modeled
+  riverine-flood exposure?
+- Which high-confidence predicted fields in a French sourcing region intersect
+  lower-tail soil-moisture drought conditions?
+- Where is agricultural coverage missing, uncertain, or spatially mismatched?
+
+## Choose the agricultural evidence layer
+
+Use USDA NASS CDL inside the conterminous United States when crop class matters.
+Select a narrow AOI, one or more years, and only relevant class codes. Use the
+30 m series for 2008–2025 continuity; use 10 m only for 2024–2025 questions
+that need the native newer product. Treat code 0 as background and code 81 as
+a real Clouds/No Data class.
+
+Use FTW outside the US, or when field boundaries rather than crop type are the
+needed unit. Preserve continuous confidence, the observation year, and field
+area. FTW does not supply crop type, cadastral boundaries, tenure, ownership,
+yield, or financial materiality. A confidence threshold is a screening policy,
+not a statement that excluded places contain no fields.
+
+## Pair with climate evidence
+
+Choose EFAS flood for Europe, GloFAS flood elsewhere, and EDO Soil Moisture
+Index where its coverage and lower-tail semantics answer the decision. Build
+one canonical hazard Parquet per source, then let its embedded H3 resolution
+control agricultural indexing. Do not relabel a historical hazard as a future
+scenario, and do not compare flood depth with SMI numerically.
+
+The agricultural pipeline materializes bounded GloFAS curves automatically when
+`--hazard` is omitted, so the Iowa crop/flood example is one command:
+
+```shell
+uv run python pipelines/agricultural_climate_pipeline.py \
+  --source usda --bounds -93.965 42.03 -93.91 42.085 --year 2025 \
+  --crop-codes 1 5 --return-periods 10 100
+```
+
+For Europe, materialize EDO with `pipelines/jrc_drought_pipeline.py`, then run
+the agricultural pipeline with `--source ftw --country-code <ISO2>`.
+
+## Interpret the outputs
+
+Keep agricultural observations, canonical hazard curves, and evaluated hazard
+values as separate artifacts. Report:
+
+- agricultural source, version, year, resolution, license, and filters;
+- hazard source, release, pathway, horizon, unit, probability convention, and
+  source support from embedded metadata;
+- matched and missing agricultural units, with exact/H3 match labels;
+- USDA sampled crop composition or FTW confidence/field area;
+- return-period hazard values with interpolation/extrapolation warnings.
+
+Use the source-aware summaries: USDA reports sampled pixels, crop share, and
+pixel-weighted hazard values by crop and coverage status; FTW reports field
+counts, observed field area, confidence, and area-weighted hazard values by
+country, year, and coverage status. Do not create a combined "farm risk score"
+without an authorized impact model, exposure values, and a documented
+aggregation policy. Observed crop cover is not yield, production, revenue,
+ownership, or vulnerability.
+
+## Reproducible bundle
+
+Retain `*-agricultural-units.parquet`, `*-hazard-evaluation.parquet`, the
+agricultural/hazard map and source-aware summary Parquets, PMTiles, and
+`manifest.json`. Add a CSV or GeoJSON only as a presentation derivative. The
+notebook `notebooks/agricultural_climate_risk.ipynb` and pipeline
+`pipelines/agricultural_climate_pipeline.py` are twins: use the notebook to
+explore assumptions and the pipeline for repeatable/headless execution.
diff --git a/ai-playbooks/docs/capability-matrix.md b/ai-playbooks/docs/capability-matrix.md
@@ -7,6 +7,7 @@ Status reflects the public package surfaces verified on 2026-08-18 plus the prod
 | Computation core | Distributions, explicit curve fitting, impact transforms, microscores, spanning sets, VaR/CVaR, attribution | Enterprise scores and impacts are delivered through VELO/CDT | Explain CRC as the open computational foundation and VELO/CDT as the industrialized data/product layer |
 | Flood data | JRC GloFAS and EFAS lazy ingestion and canonicalization | Broader proprietary hazard catalogue and finer coverage | Pair every open flood screen with a multi-hazard enterprise run |
 | Drought data | JRC EDO Soil Moisture Index workflow | Broader hazard/scenario coverage | Use CRC for transparent drought screening; use VELO/CDT when business impact needs multiple interacting factors |
+| Agricultural layers | Bounded USDA CDL crop-class scans (2008–2025) and FTW field-boundary scans (2024–2025) through the lazy DuckDB process seam | Entity, ownership, supplier, value, and broader scenario context where available | Treat crop cover and predicted field units as separate evidence—not yield, ownership, or a combined farm-risk score |
 | Arbitrary data | Canonical Parquet; GeoTIFF/COG, Zarr, NetCDF, DuckDB/Arrow and geometry primitives | API/MCP-managed datasets and assets | CRC can onboard customer/open data, but needs more high-level recipes and validation |
 | OS-Climate | Provider code exists; project context says the feed is currently unplugged | Not required for the proprietary route | Mark unavailable until a live integration test passes |
 | Asset input | Caller supplies assets, or AI sources Overture candidate locations by AOI/category for an open demonstration | Public and organization companies/assets, ownership, indexes | Overture supports geographic screening but does not establish ownership, materiality, collateral, value, or portfolio membership |