Research demonstrations

Watch causal evidence get challenged—and revised.

Begin with an introduction to Causal ABA or go directly to a full demonstration on the Asia benchmark. Both versions include captions, chapters and a complete transcript.

Intro video · Less than 5 minutes

An introduction to Causal ABA, then CArLA platform it in action.

A concise introduction to Causal ABA followed by an Asia benchmark session in which attributable evidence is inspected, challenged, resolved and rerun.

Full demo · 5 minutes

Asia benchmark full demonstration.

A complete session focused on inspecting data-derived facts, adopting language-model proposals, recording expert disagreement and revising the graph for the bnlearn Asia benchmark.

Research-use notice. The Asia network is used as a research benchmark. These demonstrations are not clinical decision-support tools.

Intro video transcript

Accessible transcript

What happens in the intro video, chapter by chapter.

0:00–0:09 · Title and purpose

Welcome to this demo of the Causal Argumentative Learning Assistant: a tool for transparent and contestable causal discovery.

0:09–0:20 · Attributable inputs

Data-derived conditional-independence tests, language-model proposals, and expert judgements enter CArLA as attributable, potentially defeasible inputs.

0:20–0:31 · Causal ABA

Causal ABA constructs arguments for causal hypotheses and exposes attacks between incompatible claims, including what must be released if a claim is forced in.

0:31–0:40 · Inspectable outputs

Its outputs are candidate causal graphs paired with an evidence ledger of retained facts, assumptions, and unresolved preferences.

0:40–0:48 · Contestability loop

Reviewers challenge evidence, resolve conflicts, and rerun. Every revision is logged.

0:48–1:08 · Load the Asia benchmark run

We now load an existing run of Causal ABA on the bnlearn Asia benchmark. Its variables cover travel, tuberculosis, smoking, lung cancer, bronchitis, X-ray results, and dyspnoea, meaning shortness of breath. The displayed graph is a candidate open to review.

1:08–1:23 · Inspect provenance

The run summary preserves the algorithm configuration and provenance behind this particular graph, rather than presenting it as a free-standing picture. Explanatory inspection makes the grounds for this candidate transparent.

1:23–1:49 · Inspect the CI ledger

The conditional-independence panel exposes the statistical evidence used by Causal ABA. Its 163 statements split into 86 retained facts and 77 defeasible assumptions. In and out answer a different question: whether each statement agrees with the displayed graph. There are 127 in and 36 out; expanding out confirms that all 36 are assumptions.

1:49–2:15 · Inspect and adopt one language-model input

Next, the language-model panel exposes structural proposals without applying them automatically. Turning on the overlay and selecting lung to X-ray brings its local evidence into view. Five of five prompts support that direction. We adopt this one proposal: it is provisionally accepted as a fact, subject to coherence with every active input, and its source remains attached for later review.

2:15–2:41 · Add one clinical input

A clinical reviewer then enters the review and challenges the candidate graph. Bronchitis can cause shortness of breath, so the reviewer proposes bronchitis to dyspnoea. Saving records a hard, attributable input for the next evaluation. It does not draw an edge directly onto the graph. This is contestability: expert judgement becomes an attributable, checkable commitment.

2:41–3:01 · Inspect the combined contestation log

The contestation log includes all prior contestations: the language-model adoption, the clinical challenge, and curator constraints on temporal scope and soft preferences. Scrolling reveals their sources, confidence, and rationales. Show all constraints visualises the complete recorded set.

3:01–3:30 · Check the resulting conflict

Now we check conflicts. One hard blocker is found: lung to X-ray is required by the adopted language-model consensus, but forbidden by a broader curator rule. The red resolution panel places the two histories side by side. The conflict is therefore an explicit attack between attributable claims, not a hidden validation failure or a silently overwritten input. Dynamic consistency checking exposes conflicting commitments.

3:30–3:46 · Resolve with a recorded exception

The review chair keeps lung to X-ray as a named exception to the broader scope rule and records the rationale. CArLA changes the formal constraint set, preserves the defeated alternative in the log, and confirms that the remaining hard inputs are coherent.

3:46–4:11 · Rerun Causal ABA

We rerun Causal ABA on the data evidence, the adopted proposal, the clinical edge, the scoped prohibitions, and the explicit resolution together. The output is derived from this exact, fingerprinted input set. When computation finishes, the revised candidate and its audit summary appear together. This is auditable redress: revisions follow from resolved inputs.

4:11–4:33 · Compare graph revisions and residual disagreement

Immediate changes distinguish additions from removals. The rerun adds bronchitis to dyspnoea and lung to X-ray, and removes dyspnoea to tuberculosis. An unsatisfied soft preference also remains visible in the summary. The system records residual disagreement for possible later contestation instead of claiming automatic obedience.

4:33–4:51 · Compare the supporting evidence

The revision releases 15 lower-ranked statements, leaving 71 facts and 92 assumptions. CArLA combines inspectable evidence and explicit redress in a user-friendly platform with an auditable workflow, enabling contestable causal discovery for high-stakes decision-making.

4:51–4:59 · Thanks and invitation

Thanks for watching. Feel free to get in touch should you have any questions or thoughts!

Full demo transcript

Accessible transcript

What happens in the full demo, chapter by chapter.

0:00–0:32 · Data and provisional graph

This is the Asia benchmark. Its eight binary variables describe recent travel to Asia, tuberculosis, smoking, lung cancer, bronchitis, a constructed either-disease indicator, X-ray results, and dyspnoea, or shortness of breath. This is Causal ABA's candidate causal graph. Node details provide variable meanings; the run summary records the data, settings, and provenance. The graph is provisional: a set of hypotheses open to challenge and revision.

0:32–1:05 · Causal ABA and the conditional-independence ledger

Causal ABA begins with 163 conditional-independence statements derived from the data. Eighty-seven currently survive the argumentative fact loop as facts, while 76 are released as assumptions. In and out answer a different question: whether the displayed graph satisfies each statement. Here 127 are in and 36 are out. Both groups remain inspectable rather than being hidden behind the summary. The interface keeps those two classifications separate.

1:05–1:50 · Inspect and adopt language-model proposals

The language-model panel contains proposals, not edits. Focusing bronchitis and dyspnoea makes their local alternatives legible. The bronchitis-to-dyspnoea direction has five-out-of-five agreement, but remains an assumption until the reviewer adopts it. The green state records that provisional acceptance. We then focus smoking and bronchitis. Clicking their arrow joins the language-model judgement with the pair's evidence. The upper rows are facts; the final lung-conditioned statement is an assumption that is out. We inspect without contesting, then adopt smoking-to-bronchitis. Adoption means provisionally accepting the assumption as fact, subject to checking full coherence.

1:50–2:15 · Add a competing expert judgement

A clinical reviewer adds a different, mechanism-based input: a bronchitis diagnosis can change subsequent smoking behaviour. That later-time influence is plausible, but it cannot coexist with smoking-to-bronchitis in one static acyclic graph. We record the proposal at full confidence and deliberately defer the conflict to Check. Once again, the old graph is not manually redrawn.

2:15–2:50 · Challenge tuberculosis directions

The clinical reviewer now rejects the displayed dyspnoea-to-tuberculosis direction. Dyspnoea is a downstream symptom and cannot cause the bacterial infection, so this is a hard prohibition and the direct disagreement appears in red. A broader, softer rule says that alternative non-Asia parents of tuberculosis remain under clinical contestation. It is guidance rather than an absolute veto. These rules constrain the next evaluation; they do not silently rewrite the current graph. The confidence value makes that difference explicit.

2:50–3:02 · Joint contestation log

Show all constraints exposes the scope. The log combines language-model and clinical-review inputs, each with its source, confidence and rationale.

3:02–3:43 · Expose and resolve the conflict

Check now finds a real blocker. Three hard required edges include both smoking-to-bronchitis and bronchitis-to-smoking, which form a directed cycle. The red resolution panel does not silently privilege either source: it places the adopted language-model direction and the clinical reviewer's later-time argument side by side. The review chair fixes the temporal scope. Smoking is treated as pre-diagnosis exposure in this static graph, while the reverse influence is retained as a softer later-time preference. The green result confirms that the hard inputs are now jointly coherent. The weaker claim remains visible for later review.

3:43–4:02 · Principled Causal ABA rerun

Causal ABA now recomputes over the saved data-derived statements and the resolved language-model and expert inputs. It searches again for a graph and fact loop that survive together. This is principled revision of both the causal output and its evidential basis, not a drawing command.

4:02–4:38 · Explain changes and unresolved preferences

The immediate-change view shows the result. Bronchitis-to-dyspnoea is added, and dyspnoea-to-tuberculosis is removed. The rerun also proposes lung-cancer-to-tuberculosis against the softer rule. The summary keeps two unresolved preferences visible: that clinical prohibition and the later-time bronchitis-to-smoking influence. Disagreement remains attributable. This is managed disagreement, not automatic obedience. The newly added lung-cancer-to-tuberculosis edge can itself be contested in a later iteration if new evidence emerges.

4:38–5:00 · Revised fact loop and takeaway

The evidence changes too. One statement is released, so 163 tests now divide into 86 facts and 77 assumptions. Causal ABA makes every input contestable: conflicts are resolved explicitly, while conclusions and unresolved judgements remain auditable and open to review.