emem-recall-polygon

This skill calls /v1/recall_polygon to fetch facts inside an arbitrary polygon. The responder samples every cell64 whose centre falls inside the polygon (cap: 4096 cells per request), recalls the requested band(s) at each, and returns a structured response with a per-cell receipt for verification.

When to invoke

The user defines a region rather than a point:

• "What's the average NDVI inside this watershed: [coords]?"
• "Show me precipitation across the area bounded by …"
• "Recall elevation across this admin boundary."
• The user pastes GeoJSON or a list of [lng, lat] pairs.

If the user just wants a single cell, use emem-locate-and-recall. If they want a similar-cells list rather than a polygon, use emem-find-similar.

How to invoke

The polygon is a closed ring: [[lng0, lat0], [lng1, lat1], …, [lng0, lat0]]. Coordinates are WGS-84 decimal degrees, longitude first (GeoJSON convention).

curl -sf -X POST https://emem.dev/v1/recall_polygon \
  -H 'content-type: application/json' \
  -d '{
    "polygon": [
      [77.55, 12.95],
      [77.65, 12.95],
      [77.65, 13.05],
      [77.55, 13.05],
      [77.55, 12.95]
    ],
    "bands": ["indices.ndvi", "weather.precipitation_mm"]
  }' | jq '{
    cells_returned: (.by_cell | length),
    facts_total:    (.by_cell | to_entries | map(.value.facts | length) | add),
    area_km2:       .area_km2
  }'

Picking the band

If the user asked about precipitation, use weather.precipitation_mm (current 24 h) or era5.precip (1940→ present hourly). If they asked about vegetation, use indices.ndvi (computed live from S2) or modis.ndvi_mean (16-day MODIS composite, larger pixels but global). If they asked about elevation, use copdem30m.elevation_mean (30 m, terrestrial only; gmrt.topobathy_mean for ocean depths). Full band list at https://emem.dev/v1/bands.

Computing aggregates

The response is per-cell facts; the agent does the rollup:

curl -sf -X POST https://emem.dev/v1/recall_polygon \
  -H 'content-type: application/json' \
  -d '<polygon_payload>' \
  | jq '
    [.by_cell | to_entries[].value.facts[] | select(.band == "indices.ndvi") | .value]
    | { count: length, mean: (add / length), min: min, max: max }
  '

For more sophisticated geometry (point-in-polygon with holes, intersection with admin boundaries), pull the full response and process locally.

Response shape

{
  "by_cell": {
    "defi.zb493.xoso.zcb6a": {
      "facts": [{ "band": "indices.ndvi", "value": 0.42, ... }],
      "receipt": { "request_id": "...", "fact_cids": [...], "signature": [...] },
      "bands_already_attested_at_cell": ["weather.temperature_2m", "indices.ndvi", ...]
    },
    "defi.zb493.xoso.zcb70": { /* ... */ }
  },
  "polygon_vertices": 5,
  "cells_sampled": 12,
  "facts_returned": 24,
  "area_km2": 1.2,
  "schema": "emem.recall_polygon.v1"
}

Pitfalls

• Polygon must be closed — the first and last [lng, lat] pair

must be identical. Open polygons return 400 invalid_polygon.

• Cap at 4096 cells. Very large regions return 413 polygon_too_large

with a hint about subdividing. The /humans page's lasso uses this cap as the geometric guide.

• Counter-clockwise orientation is conventional for outer rings;

most responders accept either, but the area calculation flips sign on clockwise input.

• Per-cell receipts — each cell's receipt verifies independently.

There is no aggregate receipt; verifying any one cell verifies that cell only. To audit the bulk call, iterate by_cell and pass each .receipt to the emem-verify-receipt skill.

Worked example

USER: What's the average NDVI inside this watershed:
      [[77.5,12.9],[77.7,12.9],[77.7,13.1],[77.5,13.1],[77.5,12.9]]?

CLAUDE invokes this skill:
  POST /v1/recall_polygon
    body: { polygon: [[77.5,12.9],...], bands: ["indices.ndvi"] }
    → cells_sampled: 12
    → facts_returned: 12
    → area_km2: ~5

  jq aggregation:
    → mean NDVI: 0.41

CLAUDE replies: "Across the 12 cells inside that polygon (~5 km²
in central Bengaluru), the average NDVI is 0.41 — typical for a
mixed urban core with pockets of vegetation. The lowest cell
returned 0.18 (heavily built up), the highest 0.62 (the
Cubbon Park area). All twelve facts are signed individually;
re-fetching any fact_cid returns the same bytes."