Processing Mode ?How modes workEach mode controls how many AI models participate and what pipeline stages run. Higher modes use more models and add verification steps, producing better answers but taking longer. Credits indicate relative cost.

Council Models ?Choose your councilSelect which tier of AI models will deliberate on your question. Budget uses free models. Balanced is best value. Premium uses top models. Ultra uses flagship reasoning models. Research is optimized for web research.

GPT-4o Mini Claude 3.5 Haiku Gemini 2.0 Flash

Chairman (synthesizes final answer)

✎ Customize models

Select Council Models 0 selected (min 2)

Research & Input ?Expand your council's knowledgeEnable web search for real-time information, or upload images for visual analysis. These give your council access to current data and visual context.

🌐 Web Search ?Real-Time Internet ResearchSearches the web for current information before the council deliberates. Uses Tavily AI search to find relevant, high-quality sources.

Best for: Current events, recent news, product comparisons, fact-checking, or any question where up-to-date information matters.

Example: "What are the latest AI regulations in the EU?" will search for recent news and policy updates.

Adds 5-15 seconds. Sources are shown and can be cited.

Searches the internet for current information before deliberation. Essential for news, recent events, or fact-checking.

📷 Image Analysis ?Visual UnderstandingUpload images for the council to analyze. Vision-capable models will examine the image and incorporate their observations into the deliberation.

Supported formats: JPEG, PNG, GIF, WebP (max 10MB each, up to 5 images)

Best for: Chart/graph interpretation, design feedback, image identification, document analysis, comparing visuals.

Example: Upload a screenshot of an error message and ask "What's causing this error and how do I fix it?"

Upload images for visual analysis. Council will examine and discuss the images in their deliberation.

Anti-Hallucination & Quality ?Why these matterLarge language models can confidently state incorrect information ("hallucinate"). These features add extra verification steps to catch mistakes. In benchmarks, these features produce 3x more nuanced responses with 100% complete answer rates compared to individual model calls.

Verified Mode ?Full Verification PipelineExtracts every factual claim from the synthesis, checks each one against research sources, and runs Chain-of-Verification (CoVe) on ungrounded claims. Adds a confidence score and highlights which claims are verified vs. unverified.

Example: If the council says "Python was created in 1989," the verifier checks that claim against sources and confirms or corrects it.

Adds 40-100 seconds of processing time.

Extracts and fact-checks every claim. Shows a confidence score. Catches errors like wrong dates, stats, or attributions.

Devil's Advocate ?Contrarian ChallengeAssigns one council member to actively argue against the group's emerging consensus. This model's job is to poke holes, find edge cases, and present the strongest possible counterargument.

Example: If 2 models agree "remote work is more productive," the devil's advocate will argue why office work can be better, forcing the synthesis to address both sides.

Prevents groupthink and echo-chamber answers.

One model deliberately challenges the consensus. Prevents one-sided answers on controversial or nuanced topics.

Include Citations ?Inline Source ReferencesAdds numbered citations [1][2][3] throughout the synthesized answer, with a reference list at the bottom showing exactly which source supports each claim.

Example output: "Python is the most popular language for AI [1], though Julia is gaining traction in scientific computing [2]."

Requires Verified Mode to be enabled. Works best with Research or Deep Analysis modes where web sources are available.

Adds [1][2] inline references to the final answer with a source list. Shows exactly where each claim comes from.

Semantic Check ?Hallucination Circuit BreakerAfter all models respond, this compares their answers semantically. If responses diverge significantly (e.g., one model says "yes" while another says "no"), you'll see a warning banner.

Example: Q: "Is coffee good for you?" Model A says "yes, in moderation." Model B says "no, it causes anxiety." The semantic check flags this disagreement so you know the answer is contested, not settled fact.

High divergence often means the question is ambiguous or the topic is genuinely debated.

Detects when models fundamentally disagree. Warns you if the answer is contested rather than consensus.

STORM Report ?Stanford STORM Research MethodologyGenerates a structured, academic-style research report based on Stanford's STORM framework. Creates multi-perspective research questions, conducts grounded expert interviews across models, and produces a formatted report with:

- Executive Summary
- Methodology
- Key Findings
- Limitations
- References

Best for: "Write a research brief on the impact of AI on healthcare" or "Compare EV battery technologies."

Requires Research or Deep Analysis mode. Adds 75-150 seconds.

Produces a structured research report (Executive Summary, Findings, Limitations, References). Uses Stanford's STORM methodology.

Iterative Refinement ?Revision Based on Peer ReviewAfter the critique stage, models that received low peer-review scores are asked to revise their response incorporating the feedback before the final synthesis.

Example: If Model A's response scores 3/10 from peers because it missed economic factors, Model A rewrites its answer to include economic analysis. The chairman then has better material to synthesize from.

Improves quality of the synthesis when initial responses are uneven. Adds one extra round of model calls.

Models that score poorly in peer review revise their answer before synthesis. Improves weak responses.

Mixture of Agents ?MoA Pre-Analysis LayerBefore the main council runs, a separate set of expert models analyze the question from different angles (technical, practical, ethical, etc.). Their analysis is fed as context to the main council models.

Example: For "Should we adopt Kubernetes?", the MoA layer might produce: a DevOps perspective on ops complexity, a cost analysis, a security review, and a migration risk assessment. The council then has richer context for deliberation.

Auto-enabled in Deep Analysis mode. Can be manually enabled for Council or Research modes for extra depth.

Expert models pre-analyze the question from multiple angles (technical, ethical, practical) before the council deliberates. Auto-enabled in Deep Analysis.

Early Stopping ?Skip Remaining Models on ConsensusIf a strong consensus emerges early in the council deliberation, stop calling additional models and proceed directly to synthesis. Saves time and cost when the answer is clear-cut.

Example: If 2 of 3 models give nearly identical answers with high confidence, the third model may be skipped.

Best for: factual questions with clear answers. Not recommended for nuanced or controversial topics where diverse perspectives matter.

Skip remaining models if strong consensus emerges early. Saves time and cost on clear-cut questions.

Attachments ?Upload Context FilesAttach documents or images for the council to reference when answering. PDFs are extracted as text; images are analyzed with vision models.

Examples:
- Upload a contract PDF and ask "What are the key risks in this agreement?"
- Upload a chart image and ask "What trends do you see in this data?"

Up to 5 files, 10MB each. Supported: PDF, JPG, PNG, GIF, WebP.

Drop files here or click to browse
PDF, JPG, PNG, GIF, WebP — up to 5 files, 10 MB each

Estimated Impact

based on mode

Time

~8s

Cost

$0.003

Confidence

Standard

Active: Single model, fast response

⚖

The Deliberation Chamber

Ask any question and watch in real time as multiple AI models independently respond, critique each other's answers, and synthesize the strongest possible answer.

How it works

1. Models Respond — 3+ AI models (GPT-4o, Claude, Gemini) each write an independent answer to your question.

2. Peer Critique — Each model reviews and scores the others' answers, identifying gaps, errors, and strengths.

3. Chairman Synthesis — A senior model reads all responses + critiques and writes the final, balanced answer combining the best insights.

In benchmarks, individual models deflect or truncate 25% of complex questions. Qracle's council delivers a complete, verified answer every time — with 3x more nuance and perfect structure.

The Deliberation Chamber

Upgrade Your Council

Starter

Pro

Power User

The Deliberation Chamber

Export Proof Bundle

Upgrade Your Council

Starter

Pro

Power User