← Back

Debates 101

Pronnpt runs structured debates across 5 frontier AI models — Claude, ChatGPT, Grok, Gemini, and DeepSeek. Each member is randomly assigned a debating role (e.g. Contrarian Challenger, Systems Thinker) to ensure diverse perspectives. In the final round, they vote freely based on the full discussion. There are 8 debate formats:

Type 1Closed Predictive

The council debates outcomes from a prediction market. Each LLM ranks the candidates and assigns probability estimates, then the council's forecast is compared against the market price to find alpha.

Example: “Who will win the 2028 US Presidential Election?”
Candidates: From prediction markets (Kalshi, PredictIt, Manifold)
Output: Ranked candidates with probabilities, council vs market alpha
Type 2Motion

A yes/no question framed as a motion. Each LLM provides a score from 0 (completely disagree) to 1 (completely agree) — above .50 counts as Accept, below as Reject. After seeing all Round 1 analyses, each member independently re-scores in Round 3.

Example: “Should UHI (as proposed by Musk) be started?”
Candidates: N/A — binary yes/no
Output: Majority verdict (e.g. 4-1 Accept), per-member scores and shifts
Type 3aThreshold

The question asks about a measurable quantity — a number, percentage, price, or count. The system generates ordered range brackets and the council ranks them by likelihood.

Example: “How high will unemployment get before 2030?”
Candidates: Auto-generated numeric brackets (e.g. 4-5%, 5-6%, 6-7%)
Output: Probability distribution across bands, winning bracket
Type 3bTimeline

A variant of threshold for "when will X happen?" questions. The system generates date range brackets and the council forecasts which time period is most likely.

Example: “When will AGI be achieved?”
Candidates: Auto-generated date ranges (e.g. 2030-2034, 2035-2039)
Output: Temporal probability distribution, winning time period
Type 3cPrice Forecast

A financial price forecast market — "when will asset X reach price Y?" The council estimates cumulative first-passage probabilities for each time cutoff. Uses mathematical framing (weakly increasing, not summing to 100%) so LLMs don't assume the event is guaranteed.

Example: “When will Bitcoin cross $100k again?”
Candidates: From prediction market time cutoffs (e.g. Before May 2026, Before Jan 2027)
Output: Cumulative probability curve, council vs market comparison with price history chart
Type 4Open Advisory

An open-ended question where each LLM proposes its own top 3 candidates. Proposals are then ranked, clustered, and aggregated into a council recommendation.

Example: “Best programming language for beginners?”
Candidates: Each LLM proposes their own picks
Output: Ranked recommendation list with consensus metrics
Type 5Multi-Candidate

The system suggests a set of real-world candidates (e.g. teams, politicians, companies) which the user can edit — add, remove, or approve. The council then ranks within this approved set.

Example: “Who will win the Premier League?”
Candidates: LLM-suggested, user-approved (up to 15)
Output: Ranked list from the approved candidate set
Type 6Comparative

The question presents exactly two alternatives. Each LLM lists pros and cons for both sides, then assigns a percentage split. The council's final split is the average of all 5 members.

Example: “In parenting, what matters more — nurture or nature?”
Candidates: The 2 options from the question
Output: Per-LLM percentage splits, top arguments for each side

How is the type chosen?

When you submit a question, a classifier determines the best format automatically. Yes/no questions become Motions. Questions comparing two things (X vs Y) are offered as Comparative or Motion. Questions about measurable quantities or dates become Threshold or Timeline debates. Questions implying a real-world set of candidates (teams, people, companies) become Multi-Candidate — you can edit the candidate list before the debate starts. Everything else runs as an Open Advisory debate.