spa-l3-paris-emergency-routing

Analyst notes

Description

This task tests whether the agent can run an end-to-end routing workflow on a live-fetched OSM graph and ship the four-layer GPKG a SAMU GIS team would actually open. The hidden gotcha is the CRS: the prompt says 'the official metric coordinate system for Paris' without naming Lambert-93 (EPSG:2154), and the agent is expected to infer the national grid from regional convention rather than fall back on a generic UTM zone.

Approach

1. Pull the driveable highway network and all hospitals from Overpass for the given bbox, watching the query size so it does not time out.
2. Build a directed routing graph that respects oneway tags and weights edges by travel time (posted maxspeed where present, 30 km/h otherwise).
3. For each incident in incidents.csv, snap to the nearest network node and run a shortest-path search to find the closest hospital by network distance; export the route geometry as a LineString.
4. Compute the network distance from each incident to its three nearest hospitals and order them by rank within each incident.
5. Generate a 15-minute drive-time isochrone around every hospital (a convex or concave hull over the set of nodes reachable inside the time budget is acceptable).
6. Reproject every spatial layer to Lambert-93 (the national grid for metropolitan France) and write a single multi-layer GeoPackage with the four required layers and column names.

Pitfalls

• Picking a generic UTM zone (32631) instead of Lambert-93 still scores most of the task but costs the canonical-CRS subcheck, because the prompt asks for the official metric CRS for Paris rather than any defensible metric CRS.
• Using Euclidean (straight-line) distance instead of network distance produces hospital matches and isochrones that look superficially right but fail the distance and IoU subchecks.
• Building an undirected graph or ignoring oneway tags lets the router cut against traffic, which systematically shortens distances and skews the rank ordering.
• Forgetting to translate maxspeed into a travel-time weight produces isochrones sized by raw distance, which collapses them on dense networks and inflates them on sparse ones.
• Writing GeoJSON or a separate file per layer instead of a single multi-layer GPKG fails Gate 1 because the four required layer names cannot be located.
• Losing CRS metadata on any spatial layer makes the file fail the format gate, and storing degree coordinates while declaring a projected CRS trips the coordinate-range sanity subcheck.

Map

Recent runs task v2