Probed 2026-09-14 against the live staging deployment. Every result below is reproducible with curl or with probe/undc.py.

Surface

• MCP endpoint: https://unsd-datacommons.gcp.un-icc.cloud/mcp — live, no auth, streamable HTTP, answers stateless JSON-RPC. Server DC MCP Server v1.3.0, protocol 2025-06-18.
• 6 tools: search_indicators, search_child_indicators, get_variable_metadata, get_observations, get_child_observations, get_multi_entity_observations.
• 3 playbook resources: skill://data-commons-researcher/SKILL.md, skill://data-commons-child-places-researcher/SKILL.md, skill://data-commons-multi-entity-researcher/SKILL.md. The server's instructions mandate reading the relevant playbook before tool calls, mandate per-datapoint attribution, and forbid guessing DCIDs.
• REST: https://unsd-datacommons.gcp.un-icc.cloud/core/api/v2/node — structural walks only.

Finding 1 — the graph is national-level. NYC has no UN data.

search_indicators(query="road traffic deaths", places=["New York City","United States"]):

• NYC resolves as an entity: geoId/3651000, type City.
• Every returned variable lists placesWithData: ["country/USA"]. NYC appears in none.
• Scoping a search to NYC alone returns zero variables and zero topics — even for "total population".
• search_child_indicators(parent="United States", children=[NYC, LA, Chicago]) for PM2.5: empty.

places is a data-availability filter, not a hint. Scope to a city and you get an empty set.

Implication: there is no "look up NYC in UN Data Commons". The crosswalk is the product.

Partial exception, worth exploring: several variables carry an URBANIZATION--DOU_CITY dimension (undata/sdg/EN_ATM_PM25.URBANIZATION--DOU_CITY, undata/unicef/DM_BU_PC_DOU.URBANIZATION--DOU_CITY). That is a national figure sliced by degree of urbanization — "the city parts of the US" — not a figure for any particular city. It may still be the most honest available comparator for a NYC number.

Finding 2 — MCP is governed; REST is federated. Discover via MCP only.

Every variable returned by MCP search is undata/-prefixed (undata/sdg/*, undata/who/*, undata/unicef/*, undata/ilo/*, undata/unodc/*, undata/itu/*, undata/unido/*). No leakage observed across any probe.

The REST guide warns that the same endpoints answer for other publishers' variables without warning — so REST can silently return a non-UN number that is structurally identical to a UN one. Rule: discover through MCP; use REST only for structural walks starting from an undata/ identifier.

Finding 3 — series density varies wildly. This can kill a demo.

Verified with date:"all":

date:"all" works correctly — UNICEF proves it. The sparsity is genuine: WHO road-death figures are periodic modelled estimates, not an annual administrative series. A "NYC trend vs SDG target" chart on 3.6.1 is not possible from this source.

This finding is what the coverage probe exists to systematize.

Finding 4 — provenance is first-class

Every observation carries provenanceUrl, a unit DCID (e.g. undata/UNIT_MEASURE-RATIO_COUNT_PER_100000_COUNT_POP), observationPeriod (P1Y) and sourceId. Unit and period are machine-readable, so a NYC per-100k rate against a UN per-100k rate is a checkable match rather than a judgment call. That makes part of our comparability grader automatic.

Consequence for the application we submitted

The pitch survives — the bridge is the value, and the platform's own governance boundary gives our comparability grade a machine-checkable basis. One promise needs revisiting: we said we would compare NYC to peer cities (London, Bogotá). UN Data Commons alone cannot do that. Either source peer-city data elsewhere (OECD metro, Eurostat Urban Audit, UN-Habitat) or reframe as NYC-vs-nations.