Ground Truth: heat research tested against New York City's own data
Each Claim is one relationship from published research (System Public Graph) tested against NYC data on
Underlay: confirmed, contested, bridged or inconclusive, with what data would settle it.
Each DistrictNeed scores one community district for how much new heat research it needs.
Every record's inputs cites the exact collection, version id and record ids it rests on.
Whose evidence: 767 heat-and-health studies in the graph; 4 are about New York City; 5 name low-income people; 3 name Black or Hispanic people; 66 are animal studies.
Hand-check
We checked 27 findings against their papers: 21 held up, 1 was wrong in the graph and excluded, 5 were unclear and kept.
- Excluded zcM5Dx94UD (bridge-trees-heat-ed): "air temperature" and "community-acquired pneumonia hospitalizations among those aged ≥ 65 years", relative risk 1.89 [1.01 to 3.56]. The paper's RR 1.89 is for low temperature (cold, lags 0-5 days); the graph dropped the qualifier, so the direction is reversed.
- Our code read "poor mental health" as the absence of harm, so 2 cold-home findings counted backwards. Fixed: the cold-harm end no longer flips on words like "poor" or "lower".
Inputs
| Collection | Version | Id |
|---|---|---|
nyc/environment-health |
v1.0.0 | ulv3:cd96b9507b2f4e362ac4d7dbbf40a8c00ef01cc7787bb7d5d7758951e3c2352e |
nyc/geographies |
v1.0.0 | ulv3:5406699a873a2f9ec89abbfa691a631c6876b1826eed5605709e31d66ca0950c |
nyc/street-trees |
v1.0.0 | ulv3:7cd6d86ebcfbe1d514139bf8d6650885ec34c82816d6bfdb742aab5d0bb0477d |
nyc/heat-complaints |
v1.0.0 | ulv3:4ce37f3ea3f54c5e0fb80267ef04026d6da61b324d851aee92aa6311c8312f48 |
noaa/nyc-daily-weather |
v1.0.0 | ulv3:4ff0c069903f66a9d1de596658d54745f1487ad74c5178b6a1213dcd18ede656 |
system/public-graph |
v1.0.0 | ulv3:cf73c4f9d7452909fa2b796791feac01a99de09eeb21b277744b0e5e96f272e0 |
Rules
- Research direction: majority sign of findings with a readable direction (ratio CI wholly above or below 1; signed coefficient or correlation with CI excluding 0 or p < 0.05). 'mixed' if fewer than 70% agree. Variables named 'lack of', 'low', 'no' etc. count in the opposite direction. Animal studies excluded.
- City test: Spearman rank correlation across areas; partial Spearman controlling for poverty; bootstrap 95% CI; Benjamini–Hochberg false discovery rate 10% across all city tests.
- Confirmed: city sign matches research, significant after correction, and the sign survives poverty control.
- Contested: city sign opposes research and is significant, or opposes it both before and after poverty control.
- Bridged: research links each step of a chain, no single dataset holds both ends, and the city end-to-end test has the predicted sign and is significant. Each middle link is reported.
- Inconclusive: everything else. Every non-confirmed claim says what data would settle it.
Code: https://github.com/alighazi288/system-public-challenge-2026/tree/main/projects/neural-forge at commit 67004d192f6c86455df561ae7194ef2b6b2117ab.
Built for the System Public Challenge 2026, Build Challenge 4. Derived from the System Public Graph (System, Inc. (2026). System Public Graph (Version 2.1) [Data set]. Harvard Dataverse. https://doi.org/10.7910/DVN/7V6JDO), so this collection is CC BY-NC 4.0.