CausalMosaic reads the empirical literature, extracts every causal claim it can verify, and assembles the fragments into a single Structural Causal Model, so that domain experts can ask “what if?” and get an answer with the studies attached.
Prediction tells you what will happen. Causation tells you what to do. The ground truth we need – a complete causal system – already exists, fragmented across millions of papers, trials, and guidelines. CausalMosaic combines the literature with the user's own data: the literature provides the causal structure, the user's data steers which questions are worth asking. LLMs do structured extraction only; the math is deterministic.
Parse the literature for causal claims: variables, direction, effect size, study design, sample size.
Score each claim by methodological rigour with a GRADE-inspired framework.
Stitch the fragments into a single, consistent model. Conflicts get flagged. Uncertainty stays transparent.
Query all scenarios against the model, conditioned on your own data; with pathways and sources attached.
Describe what you’re trying to decide. CausalMosaic proposes the variables worth tracking, synthesises the literature, ranks interventions by expected effect, and shows the result as a range of outcomes and their uncertainty. Every claim collapses into the underlying studies.
Demo · the walkthrough is scripted. Both examples run on fictional studies, of invented cohorts and datasets: every number is fitted from their data, and every claim traces back to the sentence it came from. Press Play, or jump to any step.
Decision making in medicine, policy, or regulated business takes more than confidence scores from a chat box. CausalMosaic is built around six explicit guarantees, each surfaced in the interface.
Every causal claim is rooted in peer-reviewed studies. LLMs do structured extraction only; the math is done deterministically.
Hundreds of studies per relationship, each weighted by methodological quality. Wisdom of the crowd, with transparent voting.
Answers are ranges of outcomes. Every edge is drawn by its effect, its edge probability and its ensemble agreement: how strong it is, how likely it is to exist, and how far the studies agree.
A Structural Causal Model is interpretable by design. "A causes B with strength X" – no neural-network shrug.
Every claim traces back to its primary evidence. Click an edge, see the studies, open the passage. EU-AI-Act-ready by construction.
Run an intervention – the pathways that drove the answer light up. The causal structure is the explanation.
The domain expert sets the question and the room to move. Everything downstream – literature retrieval, graph synthesis, probability estimation – is auditable and deterministic.
The user describes the situation in chat and, optionally, drops in any data they already have – a patient record, a cohort summary, a business profile. From that, CausalMosaic proposes a candidate set of observed variables and the levers available to intervene on. The user accepts, edits, or extends the set before the system goes to the literature.
For each pair of variables, we pull the relevant peer-reviewed studies and score them by methodological rigour (GRADE-inspired). LLMs are kept on a short leash – they only do structured extraction of effect sizes, study designs, and sample sizes. No hallucinated findings reach the model.
Bayesian aggregation stitches the fragmented findings into a single causal graph. Each edge carries the probability that the relationship exists at all, and how much the underlying studies agree on it. Where they disagree, the edge stays flagged for the user rather than being averaged into a false consensus.
The model needs to know how each variable's outcome depends on its parents. Studies usually report only summary statistics, often of the same concept through different projections – pack-years, years-smoked, binary smoker. We reconstruct the latent distribution that produced those projections, then use it to parameterise each node of the model.
Text2Tabular – ACL 2026Verifiable Parameterisation of BNs – ACL 2026With the model in place the user asks the questions that matter: observational ("what's the risk given this profile?"), interventional ("what changes if I act on X?"), or counterfactual ("what would have happened if…?"). CausalMosaic proposes candidate interventions, deterministic algorithms score each one on the network – expected outcome, shift in distribution, remaining uncertainty – and the user sees a ranked set of options with the trade-off made explicit: low-risk-medium-impact vs. high-impact-higher-risk, with the causal pathway and supporting studies attached to every claim.
The pipeline is being built one paper at a time. Two papers are published at ACL 2026, covering data reconstruction and Bayesian-network parameterisation; further work is in active development.
A unified framework combining LLM-driven information extraction with copula-based distribution modelling and constrained MCMC refinement. Reconstructs realistic research datasets from publication summaries alone – preserving multivariate relationships of the originals on Train-on-Synthetic-Test-on-Real benchmarks.
Five strategies for reconstructing local conditional probability tables from statistical summaries alone, evaluated across mixed-type networks. Copula-based methods significantly outperform baselines – unlocking BN parameterisation when raw tabular data is unavailable.
Coverage benchmarks across PubMed, OpenAlex, Google Scholar, and similar indexes – and a preprocessing pipeline that normalises metadata across them (e.g. via the Unified Medical Language System), so the same concept can be matched reliably regardless of source.
Extracting findings from every part of a medical publication – text, tables and figures. Trial reports draw most of their figures from a few types (flow diagrams, Kaplan–Meier curves, forest plots), so their structure can be detected explicitly. Every answer comes only from what was extracted and points to its source in the document.
A clickable Decision Lab prototype tested with a first cohort of domain experts. Evaluating usability of the chat-driven setup, trust in the synthesised graph, and end-to-end uncertainty propagation from primary evidence through to the ranked interventions on the screen. The findings will steer the next research cycle.