Skip to content

Consensus

Consensus is a model-assisted evidence map over the papers available to JARVIS. It is not a statistical consensus estimate, systematic review, or meta-analysis; missing papers, extraction errors, and model interpretation can change the result.

The Consensus page (at /consensus, in the Ⅱ Read sidebar group) shows where the papers in your library agree and disagree on shared claims extracted by the contradiction-detection pipeline.

Agreement by claim

The top card is titled Agreement by claim and contains a horizontal stacked bar chart — one row per shared claim topic. Each bar is split into two segments:

  • Supports (green) — the number of cross-paper assessments where one paper supports the claim.
  • Opposes (red) — the number where a paper contradicts it.

Hovering a bar shows the exact counts.

Claim evidence

Below the chart, each claim appears as an expandable card. The card header shows the claim topic and a summary badge such as 2 support · 1 oppose.

Click Show evidence (N) to expand the card and read the individual assessments. Each assessment shows:

  • The stance (Supports or Opposes) in the corresponding color.
  • The title and quote from the first paper, with a page reference where available.
  • The title and quote from the second paper it was compared against, also with a page reference.

Click Hide evidence to collapse the card again.

Empty state

When the pipeline has not yet run a consensus scan, the page shows:

No related-paper claims yet Run a contradiction scan across related papers to see where they agree and disagree.

Click Run consensus scan to queue a full contradiction scan across your library. The button label changes to Scanning… while the job is queued or running, and the background-task menu shows progress. When the whole-library job succeeds, the Consensus page refreshes its data automatically.

Requirements

A consensus scan can only produce claims when both preconditions hold:

  1. Summarized papers with verified findings. Papers must be processed so their key findings (with verified quotes) exist. If your library has none, the scan request is skipped immediately with a "Process some papers first" message instead of queuing a job that cannot find anything.
  2. Cross-references between library papers. The library-wide scan only compares pairs of papers that were cross-referenced as related during summarization. A library of processed but unlinked papers finishes instantly and explains that none of your processed papers are cross-referenced yet.

A completed scan can still produce no chart. In that case the page explains why rather than showing a generic empty state: it distinguishes no cross-referenced candidate pairs, candidates found but none passed quote verification (with the pair counts shown beneath the empty state), and a true zero-cluster outcome. A failed scan — including the case where the analysis model was unreachable for every candidate pair — shows the failure message and leaves the retry action available. JARVIS does not lower quote-verification thresholds just to populate this view.

  • Paper Detail — the Contradictions card in the right pane triggers a per-paper contradiction scan and lists contradictions for that paper.
  • Citation Graph — explore citation relationships between the same papers.
  • Knowledge Graph — visualise the entities and relationships extracted from your library.