AERIS: Agent Traces and Benchmark for Replicating Public-Health Air-Pollution Studies

This dataset accompanies the paper "Towards Autonomous Environmental Health: AERIS, a Research Agent for Air-Pollution Exposure Assessment" (NeurIPS 2026 Workshop on GenAI for Health). It contains the full execution traces of AI coding agents autonomously replicating published public-health air-pollution studies, alongside the 18,919-record literature screening corpus and the source-substitution reanalysis benchmark.

Abstract: Air pollution harms human health and contributes to millions of deaths each year, yet many public-health studies omit it even when they record the locations and dates needed to estimate exposure. We introduce AERIS (Airborne Exposure Research & Investigation System), a research harness that gives coding agents the domain tools to link study locations and dates to public air-quality records. To evaluate AERIS on a complete scientific workflow, we introduce source-substitution reanalysis: reconstructing a published study's pollution exposure from independent public measurements, linking it to the study's health outcomes, and re-estimating the health effects. Screening 18,919 papers yielded nine published studies with open health data, reproducible statistical models, and sufficient air-quality data coverage. Across these studies, AERIS processed 1.85 million monitor observations into 20,568 area-time exposure estimates and 14 pollution–health comparisons across four pollutants. With researcher review, AERIS matched the direction of all seven published positive associations, with a median study-level difference of 0.040 between AERIS and published log effect ratios. In three independent autonomous campaigns, self-driving AERIS completed 23 of 27 reanalyses compared with 7 of 27 for the same model without the AERIS harness, although removing researcher review raised the median log-effect difference to 0.169. AERIS turns air-pollution exposure assessment into an auditable, autonomous agent workflow, and clarifies where human scientific judgment remains essential.

• Project Website & Paper: https://rlacombe.github.io/aeris/
• Code & MCP Server Repository: https://github.com/rlacombe/aeris
• Authors: Romain Lacombe, Ziv Lautman, Nadav Efrat, Yougeng Lu, and Michael P. Snyder (Stanford University & University of Illinois Urbana-Champaign)

Quick Start

from datasets import load_dataset

# Load the 17 primary exposure-health effect comparisons across 9 studies
effects = load_dataset("rlacombe/aeris", "effect_comparisons", split="test")

# Load the 55 clean-room control run summaries (Self-Driving AERIS + Raw-Agent)
controls = load_dataset("rlacombe/aeris", "agent_control_runs", split="test")

# Load the 18,919-record PubMed/OpenAlex literature screening ledger
screening = load_dataset("rlacombe/aeris", "pubmed_screening_ledger_18919", split="test")

Dataset Summary

AERIS (Airborne Exposure Research & Investigation System) equips command-line coding agents with a Model Context Protocol (MCP) server, an ambient exposure assessment engine, and a durable workspace for epidemiological reanalysis.

This dataset provides the complete empirical records from the paper across three components:

1. Source-Substitution Reanalysis Benchmark (data/):

   • effect_comparisons: 17 primary exposure–health effect contrasts across 9 completed same-pollutant reanalyses comparing published odds/rate ratios against AERIS reconstructions (median ratio factor 1.04×, log-effect correlation r = 0.834, 16/17 same direction).
   • study_benchmark: Study-level metadata and execution outcomes across all 14 runnable candidates (15 execution attempts; 9 fitted reanalyses, 6 documented coverage/access blockers).
   • agent_control_runs, self_driving_study_results, self_driving_effect_comparisons: Summary tables across all 27 zero-intervention Self-Driving AERIS control runs (3 campaigns × 9 studies; 27/27 completed, 8/9 fitted per campaign) and all 27 matched Raw-Agent ablation runs (3 campaigns × 9 studies).
   • provider_attribution: Station- and provider-level attribution and license audit (OPENAQ_ATTRIBUTION_MANIFEST.csv) for all nine fitted reanalyses.

2. Literature Screening Corpus (screening/):

   • pubmed_screening_ledger_18919: The frozen 18,919-record PubMed/OpenAlex screening ledger and title/abstract classifications.
   • route1_strict_and_rescue_ledger: Route 1 strict and rescue screening classifications (1,405 candidate records).
   • route2_named_source_audit_202: Route 2 full-text and source audit of 203 named public-health outcome systems.
   • candidate_21_feasibility_audit: Detailed agent feasibility review across the 21 deduplicated candidate studies, yielding 14 runnable studies.

3. Complete Clean-Room Agent Trajectories (trajectories/):

   • trajectories/self-driving/: Full container run registries, run_manifest.json files, result.json outputs, generated analysis scripts, and non-interactive transcript-attempt-*.jsonl logs for all three Self-Driving AERIS campaigns (r8404, r8406, r8407).
   • trajectories/raw-agents/: Full run_manifest.json files, result.json outputs, generated scripts, and transcript.jsonl logs for all 27 matched Raw-Agent ablation runs (r1, r2, r3).
   • Every control run was executed non-interactively via codex exec --json inside an isolated container with zero human scientific intervention (human_scientific_interventions = 0).

Data Licensing & Upstream Attribution Notice

To respect upstream monitoring provider terms, this dataset does not redistribute raw air-quality station observation tables. Instead, it provides the complete screening ledgers, statistical comparison tables, SHA-256 run manifests, non-interactive agent trajectories, and executable retrieval scripts (in rlacombe/aeris) that reconstruct exposures directly from public sources.

We thank OpenAQ (ROR: https://ror.org/03a79pm79) for maintaining open programmatic access to global air-quality observations, as well as the national, regional, and municipal monitoring agencies and sensor networks worldwide whose measurements make exposure reconstruction possible (including U.S. EPA AirNow, the European Environment Agency, Arpae Emilia-Romagna, ARPA Lazio, CETESB, IEMA, and PurpleAir, audited in data/provider_attribution.csv).

Citation

@inproceedings{lacombe2026aeris,
  title     = {Towards Autonomous Environmental Health: AERIS, a Research Agent for Air-Pollution Exposure Assessment},
  author    = {Lacombe, Romain and Lautman, Ziv and Efrat, Nadav and Lu, Yougeng and Snyder, Michael P.},
  booktitle = {NeurIPS 2026 Workshop on GenAI for Health (GenAI4Health)},
  year      = {2026},
  url       = {https://github.com/rlacombe/aeris}
}