For the UN System Data Commons platform team. Prepared 16 Sep 2026, widened 18 Sep from the SDG goal framework to the whole governed graph — which more than doubled the surface and added two errors. By the NYC Voluntary Local Review team (Builders' Day participants), offered constructively — we are building a city↔UN indicator crosswalk and these surfaced while checking the data our own work depends on.

How they were found. A sweep of the governed corpus — 1,661 indicators, 1,701,211 observations, every reporting country and every year — against plausibility checks that need no subject-matter knowledge: a percentage outside 0–100, a negative count, a rate exceeding its own denominator, a value far outside its own indicator's distribution. One get_child_observations(variable, Earth, Country, date="all") call per indicator.

Source: probe/smell.py · full output: smell test artifact

Every item below was checked by hand against the indicator's own distribution before being included. The sweep produced 3,844 findings; these are the seven we are confident are errors, plus one that is not an error and matters more. Issues we investigated and dismissed are listed at the end, so you can see what the checks get wrong.

Summary

1. Kyrgyzstan: safety percentage is 100× too large

undata/sdg/VC_SNS_WALN_DRK — Proportion of population that feel safe walking alone around the area they live after dark. Unit Percent. Provenance unstats.un.org/sdgs/dataportal, sourceId 3722936608695937273.

The indicator holds 258 observations across 56 countries. Every other observation falls between 22.8 and 95.0, median 72.0.

Mechanism: dividing the three values by 100 gives 67.1, 68.4, 69.9 — which continues Kyrgyzstan's own trend from 66.8 (2020) smoothly. Consistent with a submission in basis points, or a proportion multiplied by 10,000 rather than 100.

Suggested correction: 67.10, 68.40, 69.90.

2. South Africa: municipal waste recycled exceeds world output

undata/sdg/EN_MWT_RCYV — Municipal waste recycled. Unit WEIGHT_TN (tonnes). sourceId 11758492570122983502.

Global municipal solid waste generation is on the order of 2×10⁹ tonnes per year. The 2019 value alone would be more than 1.5× all municipal waste generated on Earth. Across the other 114 countries the maximum ever recorded is 6.27×10⁷ tonnes, and the median is 444,000.

Mechanism: the values read as kilograms. 1.86×10⁹ kg = 1,855 kt, which against South Africa's own 2006 figure of 521 kt is plausible growth over twelve years.

Suggested check: whether the 2018– series changed submission units from the 2005–06 series.

3. Brunei: national total placed in a per-capita field

undata/sdg/EN_HAZ_PCAP — Hazardous waste generated, per capita. Unit WEIGHT_KG. sourceId 6461446536147282263.

Brunei, every year 2016–2023: 8.68×10⁶ to 3.61×10⁷ kg per person. Across the other 116 countries the median is 22.0 kg and the maximum is 211,720.

At 3.61×10⁷ kg per capita and a population near 450,000, the implied national total is 16 billion tonnes of hazardous waste per year.

Mechanism — this one identifies itself. Divide the reported value by Brunei's population:

1.2575×10⁷ kg ÷ ~450,000 people ≈ 28 kg per capita

against a global median of 22.0. The national total, in kilograms, has been written into the per-capita field. Every year from 2016 to 2023 behaves the same way.

4. Guadeloupe: one 2022 unit slip, propagated through four indicators

All four share sourceIds 6461446536147282263 (per-capita) and 11758492570122983502 (totals).

Guadeloupe's per-capita series is otherwise smooth and unremarkable across fourteen years: 1.91, 3.94, 6.33, 7.20, 7.23, 7.32, 8.46, 8.54, 10.03, 10.36, 10.58, 11.97, 13.10, 13.71.

The consequence is visible at world level. The largest total e-waste recycled ever recorded by any other country is 899,287 tonnes. Guadeloupe's 2022 figure of 5,367,000 tonnes makes a territory of roughly 380,000 people the world's largest e-waste recycler by a factor of six.

Mechanism: a ×1,000 unit slip in the 2022 submission (≈1017× on the per-capita figures, ≈981× on the totals), carried into every derived indicator.

Suggested correction: 13.95 kg per capita; 5,367 tonnes.

5. Negative remittance costs

undata/sdg/SI_RMT_COST — Average cost of sending $200 to a receiving country, as a proportion of the amount remitted. Unit Percent.

Myanmar 2022 carries −0.56 on the same indicator, and Ghana 2022 carries −4.04 on the corresponding sending-country indicator.

A cost expressed as a proportion of the amount remitted cannot be negative, and in Malawi's case the two negative years sit between 13.13 and 31.48.

6. Brazil: 1.28 billion workers involved in strikes

undata/ilo/STR_WORK_NB — Number of workers involved in strikes and lockouts. Unit COUNT_PERSONS. 1,178 observations across 91 countries.

Brazil's population is about 210 million. The 2015 figure is six times the entire population, and fourteen times the 92,324,000 maximum any country has ever recorded on this indicator. The global median is 9,831.

Brazil's own series runs between 0.8 and 3.8 million from 2000 to 2012, so 2015 is a break of roughly 400× against its own history, and 2016–2019 stay in the hundreds of millions before returning to normal.

Suggested check: whether the 2015– figures are worker-days or some cumulative measure rather than a count of persons.

7. Slovenia: unconverted tolar in an international-dollar field

undata/ilo/EAR_INEE_NB_PPP — Monthly minimum wage in international dollars at Purchasing Power Parity rates. Unit CR_USD_PPP_2021. 3,369 observations across 162 countries.

The global median is 382 and the maximum any country has ever recorded is 10,259. Slovenia's 2000–2006 values are ten to seventeen times that maximum.

The break is diagnostic. It falls exactly at 2007 — the year Slovenia adopted the euro — and every value from 2007 onward is unremarkable. The pre-2007 figures read as Slovenian tolar that were never converted into the PPP international dollars the unit declares. At roughly 240 tolar to the euro, 170,774 tolar is about 712 euro, which is the right order for the 2007 figure of 763.

Suggested check: whether other pre-euro-accession members carry the same pattern in this series.

Not an error, and more important than any of the above

undata/sdg/VC_DSR_MORT — Number of deaths due to disaster. Unit COUNT.

Across the other 163 countries the median is 42 and the all-time maximum is 222,608. The US 2020 and 2021 values track reported national COVID-19 mortality closely. The United States appears to have classified the pandemic as a disaster and reported it here; most countries did not.

This is presumably correct reporting, and that is the problem. Two numbers share a variable, a unit and an axis, and are not the same measurement. A chart of "disaster deaths, US vs peers" would be perfectly well-formed and would mislead every reader. There is nothing in the observation metadata — unit, observation period, provenance — that would let a tool detect it.

For a platform whose value rests on cross-national comparison, we think this is worth a machine-readable signal at the observation level: a note, a flag, or a method qualifier that a client can surface on the face of a chart. We would be glad to discuss it at Builders' Day; it is the single clearest case we have found for the comparability grading our project is built around.

Open questions, not claims

Three readings we cannot settle from outside the data. Each is a question for the platform team rather than a finding, because in every case the honest answer may be "that is what the number means".

What we checked and dismissed

Included so you can judge the checks' precision rather than take it on trust. Each of these looked like a finding and is not:

• Kuwait's water stress at 3,850% (ER_H2O_STRESS). Correct — withdrawal beyond renewable resources via desalination and fossil groundwater. 371 observations exceed 100% across 17 countries.
• 132,810 disaster-affected persons per 100,000 in the Marshall Islands (VC_DSR_DAFF). Correct — a person counts once per disaster, so a state hit repeatedly in one year exceeds its own population. Our check's premise was wrong, not the data.
• Euro-area countries sharing a conversion factor of 1.08271; Benin, Burkina Faso and Cameroon sharing 710.208. Correct — the check had found the euro and the CFA franc.
• Iran's monthly minimum wage, 23× the indicator's 99th percentile. Correct. The unit is CR_LCU — local currency — so 53 million is an ordinary monthly wage in rial and absurd in euro. Our outlier check compares every country's values within an indicator, which is meaningless on a local-currency series; it now skips them. Worth noting as a general hazard for anyone building automated checks on this graph.
• 23,172 negative percentages. Correct, nearly all. "Annual growth rate of real GDP per capita", "Current account balance as a proportion of GDP", "Change in minimum river flow (%)". Which leads to the one structural observation we would offer:

The Percent unit covers both bounded proportions and signed rates of change, with nothing in the unit string to tell them apart. Any automated quality check — ours or yours — has to infer the difference from the data rather than read it from the metadata. A distinct unit, or a bounded: [0,100] property, would make a large class of errors mechanically detectable. Our workaround is to treat each indicator as its own control group: a rule broken by most of an indicator's observations is its definition, and one broken by three country-years in three thousand is an error.

A third structural item

Alongside the 170 unnamed indicators and the Percent unit, one more that surfaced from widening the sweep: enumerating the corpus depends on where you start, and no entry point sees everything.

Walking ->relevantVariable from undata/topic/Root yields 1,661 base indicators; walking from the seventeen SDG goal trees yields 689. That much is expected — the goal framework is a subset. What is not expected is that six indicators are reachable from the goal trees and not from Root:

undata/sdg/SG_DSR_SILN   undata/sdg/SG_DSR_SILS   undata/sdg/SM_POP_REFG_OR
undata/sdg/VC_DSR_AGLH   undata/sdg/VC_DSR_CHLN   undata/sdg/VC_DSR_HOLH

All seventeen goal trees are direct children of Root — 17 of its 42 — so a traversal from Root should be a strict superset. Three checks rule out the obvious explanations: neither walk logged a fetch error, the goal-tree walk is exactly reproducible (re-run two days apart, identical 689), and the disagreement runs both ways — twelve undata/sdg/ indicators are reachable from Root and not from the goal trees, including SG_DMK_PARLYTH* and SE_SGE_*.

So ->relevantVariable is not transitive across these hierarchies, and a client enumerating the corpus from a single root gets a silently incomplete set with no way to detect it. We would want to know which entry point, if any, is intended to be complete.

Reproducing

git clone https://github.com/sarapis/undatacommons-nyc
python3 probe/corpus.py --roots all       # the whole graph, 1,661 indicators
python3 probe/smell.py --all --corpus probe/cache/corpus-all.json

Stdlib only. Writes docs/artifacts/smell-<date>.{json,md}; the JSON carries every finding with its DCID, place, year, value, unit and provenance. Any single item above can be checked directly:

python3 -c "
import sys; sys.path.insert(0,'probe')
from undc import Client
r = Client().call_tool('get_child_observations', {
    'variable_dcid': 'undata/sdg/VC_SNS_WALN_DRK',
    'parent_place_dcid': 'Earth', 'child_place_type': 'Country', 'date': 'all'})
print([x for x in r['data']['rows'] if x[0] == 'country/KGZ'])"

Contact: the repo is public at https://github.com/sarapis/undatacommons-nyc; we will be at Builders' Day on 22 September.