From Literature to Living DAGs: An Edge-Evidence Schema and Low-Cost Extraction of Causal Claims from Cohort Studies¶
Research proposal · Matt Sakamaki · version of November 4, 2026
Summary¶
We propose an evidence schema for causal diagrams in which every edge carries the published claims that support or contradict it, and test whether low-cost typed classification and extraction models can extract those claims well enough to keep the diagrams up to date as new studies appear.
Living directed acyclic graphs (DAGs) need edges annotated with levels of evidence and revised over time, but no standard exists for recording that evidence, and current extraction resources are not built for it. PubMedCausal defines causality linguistically and drops associational, hedged and denied claims; EpiKG2DAG keeps significant associations but drops null findings. Both discard evidence that a DAG needs, including evidence for a missing arrow. We will (1) propose a three-record schema (claim, edge assertion, DAG version) with rule-based evidence grades, designed for DAGpedia, an open repository of epidemiological DAGs that we develop; (2) build a gold corpus of cardiovascular and non-communicable disease cohort studies in that schema; (3) benchmark typed decision and extraction models, which return calibrated probabilities over predefined options in a single forward pass (released from September 2026 as “System One” models), against generative large language models and fine-tuned encoders; and (4) replay the literature in publication order for two prespecified exposure–outcome pairs, LDL cholesterol and atherosclerotic cardiovascular disease as a positive control and educational attainment and atherosclerotic cardiovascular disease as a contested social pathway, to show when new evidence changes edge grades and minimal adjustment sets. To our knowledge this is the first benchmark of such models on causal extraction and the first test of automated, literature-driven living DAG updating.
Background¶
Three lines of work now meet but do not yet connect: linguistic causal relation extraction, LLM pipelines that build candidate DAGs from epidemiological literature, and a new class of low-cost typed models.
Linguistic causal relation extraction. PubMedCausal annotates 30,000 rows of PubMed abstracts and yields 6,491 adjudicated cause–effect pairs, each with full-text spans, explicit or implicit type, and intra- or inter-sentential scope. Its guideline excludes correlational, predictive and hedged statements, and asks annotators to work from the text alone, without external biomedical knowledge. On detection, PubMedBERT reaches F1 0.739 while the best prompted LLM reaches 0.409; on span extraction the best model (DeepSeek-R1-32B, few-shot) reaches a cosine pair F1 of 0.677, but only 0.05 on inter-sentential and 0.39 on implicit relations. An earlier line classifies conclusion sentences into no relationship, correlational, conditional causal and direct causal (Yu et al., 2019), reaching macro-F1 0.88.
Causal language in epidemiology. Observational researchers routinely avoid the word “causal” even when the question is causal, a practice Hernán (2018) argues makes aims ambiguous. In 1,170 articles from 18 high-profile journals, “associate” was the most common linking word in abstracts (45.7%), and action recommendations implied as much or more causality than the linking sentence in about 85% of articles (Haber et al., 2022). Wording is therefore a weak proxy for causal intent in this literature.
From literature to DAGs. Reynolds (2026) argues that DAGs should become living, shared representations, annotated with levels of evidence and revised over time; DAGpedia implements that idea as an open repository. Manual protocols such as ESC-DAGs map each study’s conclusions into a DAG and synthesise them. EpiKG2DAG automates part of this: DeepSeek-V3 extracted exposure, outcome, significance, direction and effect size from 189,266 abstracts (F1 0.920 on 150 abstracts), SapBERT normalised terms to UMLS, and graph motifs propose confounders, mediators and colliders. VarEx extracts the covariates actually adjusted for from full text with 70B-parameter models (covariate F1 0.76 for hypertension and dementia).
System One models. Since September 2026, typed decision models return calibrated probabilities over caller-defined options in one forward pass, without generating text: TypeSafe’s Jev, Liquid AI’s d1-ultra, Cloudflare’s Clef (open weights) and, for extraction of typed values, Cloudflare’s Clairic. Early independent studies report large latency and cost gains, accuracy gaps on harder tasks, workload-dependent calibration and sensitivity to option names (Tang & Zheng, 2026; Rafe & Das, 2026). In evidence synthesis, the only published uses we found are title and abstract screening.
Research gaps¶
No existing resource records, for one claim, both how the authors worded it and what epidemiological evidence stands behind it, so no pipeline can yet grade a DAG edge.
| Capability | PubMedCausal | Yu et al. 2019 | EpiKG2DAG | VarEx | This proposal |
|---|---|---|---|---|---|
| Separates causal from associational wording | Yes, excludes association | Yes, 4 levels | No | No | Yes |
| Reads causal intent behind associational wording | No | No | No | No | Yes |
| Study design and causal method per claim | No | No | Partly, design used as a search filter | Partly, design used for screening | Yes |
| Direction and effect size | Direction only, numbers removed | No | Yes | No | Yes |
| Adjustment set | No | No | No | Yes | Yes |
| Null findings kept | No | No | No, dropped | n.a. | Yes |
| Spans linked to graph nodes | No | No | Yes, UMLS | Partly, covariate categories | Yes, DAGpedia nodes |
| Model class | Encoders, open LLMs | BioBERT | DeepSeek-V3 | 70B LLMs with retrieval | System One models vs LLMs and encoders |
| Cost and latency reported | No | No | No | Time per reference | Yes, per claim |
- Wording versus intent. PubMedCausal discards associational and hedged claims, which would remove many observational findings worded as associations; its authors acknowledge that implicit causality is hard to separate from association, but treat this as annotation subjectivity, not as an epidemiological convention. EpiKG2DAG keeps every significant association as a directed edge, which removes the distinction instead. Neither records that a study adjusted for confounders, a signal that the question was causal (AJPH letter, 2018).
- Fragmented content. Claim type, direction and effect size, and adjustment set each exist in a separate tool. A DAG edge needs all three attached to the same source claim. EpiKG2DAG notes that abstracts omit full-text estimates; none of the papers names integration as a gap.
- Evidence against edges discarded. EpiKG2DAG drops 141,361 “no change” associations as not useful for causal modelling, and PubMedCausal excludes denied relations by design. In a DAG the strongest assumptions are the missing arrows, so evidence for their absence matters most; neither paper lists this as a limitation.
- No shared gold standard for epidemiological claims. EpiKG2DAG was validated by its authors on 150 abstracts; VarEx on single expert sets for a few exposure–outcome pairs; PubMedCausal reports single-run baselines. None reports agreement on claim strength or intent.
- Untested low-cost models. Every pipeline relies on large generative models. PubMedCausal suggests a cheap first-stage filter, but we found no evaluation of System One models on causal extraction as of November 2026, and the cost per extracted edge, which governs whether a DAG can be kept up to date, is unreported.
- No standard for edge-level evidence in DAG repositories. Reynolds (2026) calls for DAGs annotated with levels of evidence and revised over time, but no schema specifies how a claim becomes an edge, how claims are graded and combined, or what changes in a DAG when a new study appears.
Aims and research questions¶
The project has four aims; the schema (Aim 1) and the living-DAG test (Aim 4) are the primary contributions to DAGpedia, and the corpus and benchmark serve them.
- Aim 1. DAGpedia edge-evidence schema (gaps 1, 2, 3, 6). Propose a new schema with three record types (claim, edge assertion, DAG version), polarity and precision for null findings, and rule-based evidence grades.
- RQ1: Can two trained annotators fill the claim record with Cohen’s κ of at least 0.60 on claim strength, intent, design, direction and polarity?
- RQ2: Do rule-based edge grades agree with grades assigned independently by the second annotator?
- Aim 2. Gold corpus (gap 4). Annotate cardiovascular and non-communicable disease cohort studies in the schema.
- RQ3: What share of claims is worded as associational but coded as causal intent, and what share of evidence would PubMedCausal-style or EpiKG2DAG-style rules discard?
- Aim 3. Extraction benchmark (gap 5). Compare System One models with generative LLMs and fine-tuned encoders on the fields the schema needs.
- RQ4: How do accuracy, calibration and cost per claim differ across model classes, and can a confidence-based cascade match the strongest model at lower cost?
- Aim 4. Living DAG test (gap 6). For two prespecified exposure–outcome pairs, a positive control and a contested social pathway, add studies in order of publication and rebuild the DAG after each addition.
- RQ5: How closely do the final literature-derived DAGs match the DAGs drawn independently by the two annotators, relative to how closely those two DAGs match each other?
- RQ6: When do edge grades and minimal adjustment sets change as evidence accrues, and does the automated pipeline find the same change points as the gold annotations?
Methods¶
Each claim in an abstract is annotated on three layers, and every field is posed to models as a typed question that System One models can answer natively.
Annotation schema¶
The unit of annotation is a claim: one exposure–outcome relation reported in the Results or Conclusion of an abstract.
| Layer | Field | Values | Question type |
|---|---|---|---|
| 1 Claim (PubMedCausal-compatible) | Causal claim present | yes / no | Noul |
| 1 Claim | Cause span, effect span | text spans | Extraction |
| 1 Claim | Causality type | explicit / implicit | Choice |
| 1 Claim | Sententiality | intra / inter | Choice |
| 2 Epidemiology | Claim strength | direct causal / conditional causal / associational / no relation (after Yu et al. 2019) | Choice |
| 2 Epidemiology | Causal intent | causal question / descriptive or predictive / unclear | Choice |
| 2 Epidemiology | Study design | prospective cohort / retrospective cohort / nested case–control / other | Choice |
| 2 Epidemiology | Causal method | conventional adjustment / g-methods or MSM / MR / target trial emulation / none | Choice |
| 2 Epidemiology | Direction | increases / decreases / null | Choice |
| 2 Epidemiology | Effect measure, estimate, 95% CI | HR, OR, RR, β; numbers | Extraction |
| 2 Epidemiology | Adjustment set | list of covariates | Extraction (full text) |
| 3 DAG | Exposure and outcome nodes | DAGpedia node identifiers | Choice over candidate nodes |
| 3 DAG | Evidence grade | derived by rule from design, method and strength | Computed |
Causal intent is judged from signals the epidemiological literature treats as causal: confounder adjustment, causal methods, and action recommendations. Each claim also carries a polarity (asserted, negated, hedged); negated, hedged and null claims are kept and graded rather than dropped.
DAGpedia edge-evidence schema (proposed)¶
The schema separates what a paper says (claim), what DAGpedia concludes about one arrow (edge assertion), and the graph built from those conclusions (DAG version), so each layer can be revised without rewriting the others.
| Record | One per | Key fields | Written by |
|---|---|---|---|
| Claim | exposure–outcome statement in one paper | PMID, spans, strength, intent, design, method, polarity, direction, estimate and 95% CI, precision class, adjustment set, extractor and confidence, review status | Pipeline, then a reviewer |
| Edge assertion | ordered pair of DAGpedia nodes | status (present / absent / undetermined), evidence grade, linked claims, conflicts, last update | Grading rules applied to linked claims |
| DAG version | exposure–outcome question | nodes, included edge assertions, minimal adjustment sets (dagitty), version date, changes since the last version | Built automatically, approved by a maintainer |
Records are stored as Markdown files with YAML frontmatter and compiled to JSON with UUIDs by continuous integration, matching DAGpedia’s current content system.
# edges/<from>__<to>.md
id: <uuid>
type: edge_assertion
from: ldl-c
to: ascvd
status: present # present | absent | undetermined
grade: moderate # high | moderate | low | conflicting
claims: [<claim-uuid>, <claim-uuid>]
updated: 2026-11-04
# claims/<claim-uuid>.md
id: <uuid>
type: claim
pmid: "<pmid>"
strength: associational # direct causal | conditional causal | associational | no relation
intent: causal
design: prospective cohort
method: conventional adjustment
polarity: asserted # asserted | negated | hedged
direction: increases # increases | decreases | null
estimate: {measure: HR, value: <x>, ci: [<lo>, <hi>]}
precision: n.a. # precise null | imprecise null | n.a.
adjusted_for: [age, sex, smoking]
extracted_by: {model: <model-version>, confidence: <p>, date: <date>}
reviewed: false
A null claim counts as evidence against an edge only when its 95% CI lies within prespecified equivalence bounds (initially HR or RR 0.90 to 1.11); otherwise it is recorded as imprecise and leaves the edge status unchanged. Initial grading rules, to be refined in Aim 1 and tested against experts in RQ2:
| Grade | Rule |
|---|---|
| High | Two or more claims with consistent direction from trials or designs with strong identification (g-methods, MR, target trial emulation) |
| Moderate | Two or more adjusted cohort claims with consistent direction and causal intent |
| Low | A single claim, or claims without confounder adjustment |
| Conflicting | Precise claims in opposite directions, or precise nulls alongside positive claims |
An edge is marked absent only with two or more precise null claims and no moderate or high evidence for its presence.
Corpus¶
The corpus covers prospective cohort studies of cardiovascular disease and other non-communicable diseases published 2015–2026 and indexed in PubMed.
- Pilot (30 abstracts). Already retrieved; 6 non-cohort records (comment, reviews, meta-analysis, cross-sectional and prediction studies) are kept as negative examples. Used to draft and stress-test the guideline.
- Main set (400 abstracts). Stratified by exposure family (lipids, blood pressure, glycaemia, adiposity indices, lifestyle and diet, environment, social determinants) and by publication period.
- Full-text subset (100 articles). Open-access articles from PubMed Central, used only for the adjustment-set field.
Annotation protocol¶
Two people annotate remotely through a web annotation tool that keeps each person’s labels hidden from the other until both have finished a record. The first annotator labels every record on all fields and, like A, draws a DAG for each prespecified pair before reading that pair’s abstracts or any model output; both annotators’ DAG files are time-stamped on OSF. A second annotator (A), from outside the project, first draws DAGs for both prespecified pairs before reading any abstract or model output, then independently labels a 15-abstract calibration set and a random 30% of the main set on the core fields: claim strength, causal intent, polarity, direction and design. Disagreements are resolved by consensus discussion between the two annotators, closed before any test-set model output is seen; A also grades a subset of edges independently (RQ2). Spans and estimates are labelled by the first annotator only, and null precision is computed from the extracted confidence intervals. The guideline is frozen after the calibration round. Agreement is reported per field before consensus: Cohen’s κ for categories, token-level partial-match F1 with Hungarian alignment for spans (as in PubMedCausal), and set-level F1 for adjustment sets (as in VarEx).
Models and pipeline¶
- System One models: Jev (TypeSafe), d1-ultra (Liquid AI), Clef (Cloudflare, open weights) for Noul and Choice fields; Clairic (Cloudflare) for spans and numbers. Exact version and run date are logged for every call.
- Generative baselines: one frontier LLM and one low-cost LLM, both with structured output, zero-shot and few-shot.
- Discriminative baselines: PubMedBERT fine-tuned on the training split; the Yu et al. 2019 classifier for claim strength; an open LLM read through label log-probabilities on the same questions.
- Cascade: System One answers above a confidence threshold are accepted; the rest go to the frontier LLM. Thresholds are fixed on the development split before testing.
flowchart TD
A["Abstract<br/>or full text"] --> B{"Any claim?<br/>(Noul)"}
B -- no --> S["Skip"]
B -- yes --> C["Extract values<br/>Clairic: spans, numbers"]
C --> D["Classify epi fields<br/>Jev, d1-ultra, Clef"]
D --> E{"Confident?"}
E -- no --> F["Escalate to LLM<br/>frontier model"]
E -- yes --> G["Map to DAGpedia nodes<br/>Choice over candidate nodes"]
F --> G
G --> H["Candidate edge<br/>provenance + evidence grade"]Only claims the System One models are unsure about reach the frontier LLM; every accepted claim keeps its source and evidence grade on the DAGpedia edge.
Evaluation¶
The main set is split 100 development / 300 test abstracts, stratified by exposure family; the analysis plan is registered on OSF before the test set is touched.
- Accuracy: F1 for claim detection; macro-F1 for categorical fields; exact match for direction and estimates; span metrics as above.
- Calibration: expected calibration error and Brier score on held-out data, before and after temperature recalibration.
- Robustness: option-name swap, option-order permutation and three repeated runs per hosted model.
- Efficiency: cost per 1,000 claims and median latency, with and without caching.
- Statistics: 95% bootstrap intervals clustered by article; paired comparisons against the strongest baseline.
- External check: layer 1 is evaluated on the PubMedCausal test split once that corpus is released.
Living-DAG replay¶
For each of the two exposure–outcome pairs, all eligible claims are ordered by publication date and added one study at a time. After each addition the edge grades are recomputed and the minimal adjustment sets are recalculated with dagitty. The replay is run twice, once on gold annotations and once on pipeline output, and the two sequences of change points are compared. The two pairs are LDL cholesterol → ASCVD, a positive control with convergent evidence from cohorts, Mendelian randomisation and trials, and educational attainment → ASCVD, a contested social pathway; each has 25 abstracts sampled by publication period from a prespecified search. For each pair, both annotators draw a DAG before reading any of its abstracts or any model output. Because no correct DAG exists for the contested pair, the literature-derived DAG is compared with each annotator’s DAG and set against the agreement between the two annotators’ DAGs. Disagreements are classified by type (edge presence, variable role as confounder or mediator, unmeasured common causes) and by whether the literature can resolve them. Prespecified expectations: the LDL edge reaches moderate or high grade early and stays stable; the education edge is graded lower and changes when Mendelian randomisation or sibling designs enter; and the two annotators agree more on the LDL DAG than on the education DAG.
Analysis plan and expected outputs¶
Each research question has one prespecified primary analysis.
| RQ | Primary analysis | Decision rule |
|---|---|---|
| RQ1 Reliability | Cohen’s κ per field with 95% bootstrap CI | Fields below 0.60 are revised once, then reported as is |
| RQ2 Grading rules | Weighted κ between rule-based and expert edge grades | Rules revised on development edges only |
| RQ3 Wording vs intent | Share of claims worded as associational but coded as causal intent; share of claims, and of precise nulls, that PubMedCausal-style and EpiKG2DAG-style rules would discard; both by exposure family and causal method | Descriptive, with 95% CI |
| RQ4 Model classes | Per-field accuracy, ECE and cost per 1,000 claims; cost–accuracy frontier | A model is on the frontier if no other is both cheaper and more accurate |
| RQ4 Cascade | Macro-F1 of the cascade versus the frontier LLM alone, against share escalated | Non-inferiority margin of 2 macro-F1 points, fixed in advance |
| RQ5 DAG agreement | Edge precision and recall, and agreement of minimal adjustment sets (dagitty), for both pairs, each set against the agreement between the two annotators’ DAGs | Descriptive; disagreements reviewed with the second annotator |
| RQ6 Living updates | Number and timing of changes in edge grades and adjustment sets; agreement of change points between gold and pipeline replays | Change point counted as matched if within one added study |
Expected outputs:
- A DAGpedia edge-evidence schema (claim, edge assertion, DAG version) with grading rules, documented as an architecture decision record in the DAGpedia wiki.
- An annotation guideline with worked examples, released under an open licence.
- A gold corpus of about 430 abstracts and 100 full-text adjustment sets, released as PMIDs, spans and labels.
- Benchmark code in R and Python, with model versions, prompts and frozen thresholds.
- Two living DAGs in DAGpedia, each with its full version history and linked evidence.
- Two manuscripts: a schema and benchmark paper, and an epidemiology paper on wording, intent and living DAG updates.
Limitations and risks¶
The largest risk is that causal intent cannot be annotated reliably; the design treats that as a finding, not a failure.
| Risk | Effect | Mitigation |
|---|---|---|
| Causal intent has low agreement | RQ1 and RQ3 weaken | Anchor intent to observable signals (adjustment, method, recommendation); report κ even if low, since disagreement itself measures the ambiguity Haber et al. describe |
| Grading rules oversimplify evidence | Edge grades that experts reject | Rules fixed on development edges and tested against experts (RQ2); grades shown with their linked claims so readers can judge |
| Equivalence bounds for nulls are arbitrary | Too many or too few absent edges | Sensitivity analysis with narrower and wider bounds |
| Hosted System One models change or are nondeterministic | Results do not reproduce | Log version and date per call, repeat runs, include open-weight Clef |
| Option-name sensitivity | Inflated or deflated accuracy | Option-name swap and order tests on every Choice and Noul field |
| Abstracts omit adjustment details | Incomplete claim records | Full-text subset for adjustment sets; abstract-only fields flagged |
| Narrow domain (CVD and NCD cohorts) | Limited generalisability | Treat as a first domain; external check on PubMedCausal for the claim layer |
| Annotators are not blind to the study’s aims | Bias toward hypothesised RQ3 result | Second annotator outside the project; consensus closed before any test-set model output is seen |
| No independent adjudicator | Consensus drifts toward the lead annotator’s labels | Report pre-consensus κ as the reliability result; log whose label each consensus kept; independent adjudicator in the follow-up paper |
| Reference DAGs come from the two annotators only | A small reference set; the lead annotator is not independent of the pipeline | Both draw before any pair abstract or model output, time-stamped on OSF; report the two DAGs’ agreement as the reference, not a correct DAG; disclose the lead’s role; expert panel in the follow-up paper |
The corpus records what authors claim and how they designed their studies; it does not verify that the causal claims are true. DAGpedia edges built from it are candidate edges for expert review.
Timeline and deliverables¶
The first paper goes to medRxiv as a preprint, submitted by December 18, 2026; medRxiv screening usually takes 2–4 days and runs longer over holidays, so this leaves room to answer queries before the year ends. To fit the schedule, the preprint covers a reduced scope and the rest moves to a follow-up paper.
gantt
title Submitting to medRxiv on Dec 18 leaves room for screening before year end
dateFormat YYYY-MM-DD
axisFormat %b %d
section Setup
Schema freeze and pipeline :2026-11-05, 2026-11-10
Both annotators draw DAGs :2026-11-05, 2026-11-10
section Annotation
Pilot and guideline freeze :2026-11-11, 2026-11-17
Main annotation :2026-11-18, 2026-12-04
Design registration :milestone, 2026-11-10, 0d
Analysis-plan registration :milestone, 2026-12-04, 0d
section Analysis
Model runs, dev then test :2026-11-23, 2026-12-11
Living-DAG replay :2026-12-07, 2026-12-14
Manuscript and figures :2026-12-07, 2026-12-17
section Preprint
medRxiv submission :crit, milestone, 2026-12-18, 0d
Screening and fixes :2026-12-18, 2026-12-30Two dates cannot slip: both annotators’ DAGs for the two pairs must be drawn before any pair abstract or model output is seen (by November 10, before the calibration round), and the test set is locked at the analysis-plan registration on OSF (December 4); a short design registration, with both annotators’ DAGs attached, comes first (by November 10). Model runs on pilot and development data start during US Thanksgiving week (November 23–27), so annotation and benchmarking overlap.
| Component | medRxiv preprint (December 18) | Deferred to follow-up |
|---|---|---|
| Corpus | 150 abstracts: 100 CVD and NCD cohort abstracts plus 25 each on LDL-C → ASCVD and educational attainment → ASCVD; 30% double-annotated by A, disagreements resolved by consensus | Full 400 abstracts, full double annotation, independent adjudicator |
| Adjustment sets | From abstracts only | 100 full-text articles |
| Models | Jev, d1-ultra, one open model (MedJev or Clef), Clairic, one frontier LLM; three repeat runs | Fine-tuned encoders, label log-probability baseline |
| Living DAG | LDL-C → ASCVD (positive control) and educational attainment → ASCVD (contested), each against DAGs drawn independently by both annotators | Further pairs; independent expert panel |
| Grading rules | Applied to all edges; graded independently by A on a subset (RQ2) | Larger expert panel for RQ2 |
| Schema | Proposed and documented as a DAGpedia architecture decision record | Implemented in DAGpedia’s build pipeline |
medRxiv requires these statements at submission, so they are drafted in the first week rather than at the end:
- Ethics: the study uses only published abstracts and involves no human participants; confirm whether a non-human-subjects determination is needed.
- Competing interests covering the past 36 months, including current employment, and a funding statement.
- Data availability: corpus (PMIDs, spans, labels), code, prompts and frozen thresholds on GitHub and OSF, with links.
- Author list, affiliations and consent of every author to post.
medRxiv can later transfer the manuscript and metadata directly to partner journals, which include the American Journal of Epidemiology.
References¶
- Kunle-John I, Paul J, Agbaakin O, et al. PubMedCausal: A span-level annotated corpus for causal relation extraction in biomedical text. arXiv:2605.28363, 2026.
- Yu B, Li Y, Wang J. Detecting causal language use in science findings. EMNLP-IJCNLP 2019:4664–4674.
- Hernán MA. The C-word: scientific euphemisms do not improve causal inference from observational data. Am J Public Health 2018;108(5):616–619.
- A clarification on causal questions: we ask them more often than we realize. Letter. Am J Public Health 2018;108(8):e4.
- Haber NA, Wieten SE, Rohrer JM, et al. Causal and associational language in observational health research: a systematic evaluation. Am J Epidemiol 2022;191(12):2084–2097.
- Reynolds RJ. Living DAGs: the future of DAGs in epidemiology. Am J Epidemiol 2026;195(5):1365–1367.
- Ferguson KD, McCann M, Katikireddi SV, et al. Evidence synthesis for constructing directed acyclic graphs (ESC-DAGs). Int J Epidemiol 2020;49(1):322–329.
- Deng G, Du J. EpiKG2DAG: a framework for automated DAG construction from biomedical text. medRxiv 2026.
- Pradhan MM, Upadhayaya R, Wenyon SC, et al. A pilot project leveraging large language models for automated screening and variable extraction in observational studies. medRxiv 2026.
- Tang L, Zheng Y. Typed decision models: an early evidence audit and evaluation checklist. arXiv:2609.32160, 2026.
- Rafe A, Das S. Benchmarking System One decision models against trained classifiers and language models for automated decision gates. arXiv:2610.00346, 2026.
- Title and abstract screening for systematic reviews with Jev, a System One model. medRxiv 2026.
- Liquid AI. d1-ultra: introducing the world’s best decision model. 2026.
- Cloudflare releases Clairic for typed data extraction. 2026.