Questions & Answers

Confidence % is the Rapporteur's estimate of how well-supported each claim is, based on the discussion. It reflects model agreement — not an absolute measure of truth.

• 100% — all participating models agreed on this point without needing discussion.
• 70–99% — most models agreed, or initial disagreement was resolved during discussion.
• Below 70% — significant disagreement remained after all discussion rounds.

A high confidence score means the AI models converged — it does not guarantee the claim is factually correct. Always verify important claims independently.

Each Key Claim carries a source label showing how consensus was reached:

• consensus — all models agreed from the very first round, with no discussion needed.
• discussion_resolved — models initially disagreed, but after one or more discussion rounds they converged on the same position.
• disputed — models still disagreed after all discussion rounds completed. The positions shown under the claim are each model's final stance.

The Overall Confidence is the Rapporteur's holistic assessment of the entire answer — how much the participating models agreed overall, across all claims. It is a weighted average considering how many claims were disputed vs. agreed, and how fully the discussions were resolved.

Consensable runs a four-step process:

• Step 1 — Query: Your question is sent to all selected models simultaneously. Each answers independently.
• Step 2 — Analysis: The Rapporteur reads all responses and identifies points of consensus and disagreement.
• Step 3 — Discussion: For each disputed claim, the models engage in structured rounds. Each model sees the others' positions and can update or defend its own. Disputes run in parallel.
• Step 4 — Synthesis: The Rapporteur produces a final synthesised answer, verdict, confidence scores, and key claims.

These are safety caps that prevent the discussion from running too long or costing too much:

• Max Rounds — the maximum number of discussion rounds per disputed claim. Each round, every model gets one response.
• Max Tokens — the total token budget across all model calls. The discussion stops early if this limit is hit.
• Max Minutes — a wall-clock time limit. Useful for keeping responses fast on time-sensitive topics.

Synthesis always runs to completion regardless of caps — only the discussion phase is cut short.

This is the model that facilitates the discussion, analyses consensus and disagreements, and produces the final synthesised answer. It does not participate in the initial discussion itself — it acts as a neutral moderator and synthesiser. You can choose which model fills this role from the models you've selected.

• Rapporteur — the role name used in Synthesis and Discuss modes.
• Adjudicator — the role name used in Debate mode.

For questions about recent events, current news, or real-time data, the models' training data may be out of date. Live Web Context fetches fresh information from the web and injects it into all model prompts:

• Auto — a cheap classifier model first decides whether the question needs web search. If yes, Brave Search is used; if not, the web search is skipped.
• Brave — always searches via Brave Search, returning the top 6 results.
• Perplexity — uses Perplexity's AI-powered search to produce a pre-summarised web context.
• None — no web search; models rely entirely on their training data.

Costs are based on the number of tokens (units of text) processed by each AI model. Each model has a different per-token price. The cost shown is the total across all model calls during your query, including the discussion and synthesis steps.

The initial cost shown (marked ~) is a real-time estimate based on token counts. The final actual cost is confirmed a few seconds later from the model provider and updates automatically.

Longer questions, more complex topics with more disputed claims, more discussion rounds, and larger max-token budgets all increase cost. Using larger or premium models (like Claude Opus or GPT-5) also increases cost compared to faster mid-tier models.