At a glance · Quick start · Fields · Reading the labels · Citation

ARGUS: Evidence-Grounded Auditing of Identification Assumptions

Data release for "Evidence-Grounded Auditing of Identification Assumptions in Climate-Policy Causal Evaluations" (Zhang, Xie, Parra & Correia), to appear at ClimateNLP 2026 (EMNLP 2026 workshop).

ARGUS is a structured language-model pipeline that audits the evidence a difference-in-differences (DID) study reports for its identification assumptions. It maps each paper onto an eleven-dimension assumption–implication–evidence rubric, retrieves supporting evidence, assigns a risk level per dimension, and abstains (unknown) when relevant evidence cannot be retrieved. It does not judge whether a paper's estimated effect is correct.

Flag, not judge. Every label here is a screening signal about the evidence a paper reports, not a verdict on the paper. Please read How to read the labels before using any single cell.

At a glance

Papers audited	Rubric dimensions	Human gold cells	Planted flaws
26	11	55	11 + 33
published DID studies from eight fields of economics	one assumption → implication → evidence chain each	5 papers × 11 dimensions, two annotators	base flaws, plus a harder set of variants

Configurations

Config	Rows	Description
gold_labels (default)	55	Human-pilot labels: 5 papers × 11 dimensions, from two independent annotators under a pre-registered protocol, reconciled where they differed
corpus_results	869	Per-paper, per-dimension risk labels from three pipelines on the real-paper corpus
corpus_manifest	27	Title, year and venue of each paper in the corpus
rubric	11	The identification rubric: assumption, testable implication and expected evidence per dimension
flaw_taxonomy	11	The base flaws used for flaw injection (one per dimension)
flaw_variants	33	The harder 33-variant flaw-injection benchmark (commission and omission flaws)

Quick start

from datasets import load_dataset

gold = load_dataset("yonghongzhang/ARGUS", "gold_labels", split="train").to_pandas()
runs = load_dataset("yonghongzhang/ARGUS", "corpus_results", split="train").to_pandas()

# The 26-paper corpus reported in the paper, main system only
argus = runs[runs.counted_in_paper & (runs.pipeline == "argus_two_stage")]
print(f"abstentions: {(argus.risk == 'unknown').mean():.1%}")    # 39.5%

# ARGUS next to the reconciled human labels on the five pilot papers
order = {"low": 0, "medium": 1, "high": 2}
pilot = gold.merge(argus, on=["paper_id", "dimension"])
answered = pilot[pilot.risk != "unknown"]
more_severe = (answered.risk.map(order) > answered.gold_risk.map(order)).sum()
print(f"more severe than gold on {more_severe} of {len(answered)} answered cells")   # 25 of 33

What the corpus results look like

_{Share of the 26 papers at each risk level on each dimension, from corpus_results (keyword baseline on the left, the two-stage ARGUS pipeline on the right). Grey is an abstention.}

Fields

gold_labels

Column	Description
paper_id	Paper identifier; joins to corpus_manifest and corpus_results
dimension / dimension_name	Rubric dimension id and its readable name
gold_risk	low, medium or high
ambiguous	1 if the annotators flagged the cell as open to reasonable expert disagreement
source	agreed_independently (42 cells) or reconciled after discussion (13 cells)

corpus_results

Column	Description
paper_id, dimension	As above
pipeline	argus_two_stage (the main system), full_paper_single_pass (ablation; covers 25 of the 26 papers, with no run for paper_165) or keyword_baseline
risk	low, medium, high or unknown (abstention)
counted_in_paper	False for paper_164, a byte-identical duplicate of paper_120 that is excluded from every count in the paper. Filter on True to reproduce the 26-paper corpus

corpus_manifest

Column	Description
paper_id	As above
title_as_indexed, year_as_indexed, venue_as_indexed	Title, year and venue as recorded in the source corpus index. For paper_101 and paper_128 the index venue and year are corrupted; note gives the published reference
venue_in_table	Venue as reported in the paper's corpus table
topic	Field of economics (eight in total)
counted	yes for the 26 papers in the corpus, no for the duplicate paper_164
note	Remarks on individual papers

rubric

Column	Description
dimension / name	Dimension id and readable name
assumption	The identification assumption
implication	Its testable implication
evidence	The kinds of evidence a credible paper would report

flaw_taxonomy

Column	Description
id	Flaw identifier
target_dimension	The rubric dimension the flaw should make ARGUS flag
severity	low, medium or high
description	What the flaw is
injection_method	How it is planted into a paper
detection_signal	What an auditor should notice

flaw_variants

Column	Description
variant_id	Variant identifier
target_dimension	The rubric dimension the variant should make ARGUS flag
flaw_type	commission (a section replaced by flawed but natural prose) or omission (the supporting evidence removed)
severity	low, medium or high
injection	JSON string: the edit operation and its targets, with the replacement text for commission flaws
expected	JSON string: clean_risk_max and injected_risk_min, the bounds the audit should satisfy on the target dimension
forbidden_leakage_terms	Phrases that must not appear in the injected text, so a flaw cannot be detected from its own wording

The original YAML files are under config/, and the guideline given to the annotators under annotation/. The full annotation protocol, with its log of deviations, is in the GitHub repository: annotation/human_pilot/PROTOCOL.md. No text of the audited published articles is redistributed; papers are identified by title and venue only.

How to read the labels

Please read these caveats before using the per-paper labels:

• Screening signals, not quality judgements. A label says whether a paper reports evidence adequate to support an identification assumption. It is not a verdict on the paper's causal estimate or on research quality.
• ARGUS is systematically over-severe. Against the reconciled human labels, it assigned a higher risk level on 25 of the 33 cells it answered and a lower one on none. Weighted agreement stays low even after the pre-specified calibration rule.
• Individual cells are unreliable. The human pilot covers only 5 papers. Nothing here supports a claim about any single paper on any single dimension.
• unknown means retrieval failed, not that evidence is absent. ARGUS abstains on roughly 40% of paper–dimension assessments, often because the evidence sits in a figure that text retrieval cannot reach.

Quoting a single cell as a verdict on a published paper misuses this data.

License

Data: CC BY 4.0 (see LICENSE.md). Code (on GitHub): MIT.

Citation

@inproceedings{zhang2026argus,
  title     = {Evidence-Grounded Auditing of Identification Assumptions in
               Climate-Policy Causal Evaluations},
  author    = {Zhang, Yonghong and Xie, Yong and Parra, Isabel M. and Correia, Ricardo},
  booktitle = {ClimateNLP 2026: Workshop on Natural Language Processing Meets Climate Change},
  year      = {2026},
  note      = {To appear}
}