One claim, nine independent reads, one transparent record.
Adjudged turns contested questions into multi-model panel reviews. Every step is visible. Every dissent is preserved. Every source is named.
The pipeline, from question to published page.
Collect the claim
Submissions arrive in many forms: a viral post, a question from a reader, a news cycle. We rewrite each one in neutral, falsifiable language before anything else happens.
How a question is phrased determines what answer you can get back. We strip rhetorical loading before models see the prompt.
Curate the evidence corpus
Editors compile a source list before running the panel: government data, peer-reviewed work, primary documents, credible journalism. Models receive the same evidence, in the same form.
Without a shared corpus, model disagreement just reflects training-data differences. We test reasoning, not recall.
Run the panel - nine independent reads
Nine model families read the corpus independently. No model sees another's response. Each produces a verdict, confidence score, reasoning, and update criteria.
Sequential or cross-fed runs collapse to consensus prematurely. Parallel isolation preserves real divergence.
Synthesise consensus and dissent
We aggregate verdicts and confidence into a panel score, but the dissent is shown alongside the consensus. Refusals and incomplete responses are surfaced as data.
A 9-of-9 agreement and an 8-of-9 split tell very different stories. Both deserve to be visible.
Publish the record
Each article shows the question, panel verdict, per-model breakdown, cited sources, and what evidence would update the conclusion. Revisions stay versioned and visible.
An encyclopedia of contested claims has to be falsifiable too. Public revision history is how we earn that.
Three things that prevent panel pollution.
Multi-model panels can fail in predictable ways. These are the safeguards we built in from the start.
Geographic diversity
The panel includes models trained in the US, Europe, China, India, and the Gulf. Single-jurisdiction panels inherit single-jurisdiction blind spots.
No model self-judges
Panel members do not score each other's work. Aggregation happens through deterministic rules, not a judge model.
Refusals are recorded
If a model declines to answer or hits its output limit, that response appears in the table as a documented incomplete.
What this method can't tell you.
Multi-model panels are a tool for surfacing convergent reasoning across diverse systems. They are not a truth oracle. We're explicit about the limits.
- Panel agreement is not proof - nine models trained on overlapping corpora can be confidently wrong together.
- Our corpus is human-curated. We make source-selection choices that another editor could reasonably make differently.
- Models have known regional and political biases. The diversity of the panel reduces but does not eliminate this.
- For genuinely contested empirical questions, the right output is a wide confidence band - not a confident "no".
See every claim we've reviewed
2,847 reviewed claims with the full panel breakdown for each.
Open the index → Read deeperFull methodology
The standards, the rubrics, the version history - all written down.
Read methodology → ContributeSubmit a contested claim
Have a question you'd like the panel to review? Send it in.
Submit a claim →