TL;DR
RADAR is a multi-agent debate technique where the model plays three characters with different jobs: the Politician builds a persuasive narrative out of the facts, the Scientist hunts for missing context, and the Judge moderates and delivers the verdict. The key difference from a standard pro/con debate: the roles don't argue about what was said — one defends it, the other looks for what was left out.
The standard "find the weak spots in this argument" request works poorly: the model sees the facts and agrees, because the facts are technically true. "The company grew 40%" — true. But if that growth came after a 60% crash, it's manipulation by omission. A model in "helpful assistant" mode isn't inclined to look for what's hidden — it works with what's explicitly written.
RADAR fixes this by splitting the functions: the Politician is tasked with defending the claim, the Scientist with attacking not the claim itself but the gaps in its context. At the end, the Judge synthesizes everything and delivers one of three verdicts: true, half-true, or false. The debate runs 2-3 rounds: opening → rebuttals → closing.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: POLITICIAN — builds a narrative IN SUPPORT of the claim from the available facts
STEP 2: SCIENTIST — looks for missing context, weak links, selective framing
STEP 3: JUDGE — assesses whether the arguments are sufficient → continue or deliver a verdict?
If another round is needed:
STEP 4: POLITICIAN — responds to the Scientist's critique
STEP 5: SCIENTIST — attacks the Politician's narrative, points out omissions
STEP 6: JUDGE — final verdict: TRUE / HALF-TRUE / FALSE + explanation
All in one prompt, one LLM call.
Example in Action
Task: A founder, David Harper, publicly announced: "We launched a new product and got 50,000 registered users in the first month." You need to figure out — is this real success or just a pretty number?
Prompt:
Run a structured analysis of the claim through a debate between three roles.
**Claim:** "We launched a new product and got 50,000 registered users in the first month"
**Context:** The product is a B2C personal finance app. The company spent $40,000 on launch advertising.
---
ROUND 1. OPENING
[POLITICIAN]: Build a persuasive narrative confirming that this is a strong result. Rely only on what was said.
[SCIENTIST]: Check the same claim for missing context. What isn't said here? What data is absent? What could this wording be hiding?
[JUDGE]: Assess — are the arguments sufficient for a verdict, or is a second round needed? Answer briefly: CONTINUE or STOP.
---
ROUND 2. REBUTTALS (only if the JUDGE said CONTINUE)
[POLITICIAN]: Respond to the Scientist's critique. Explain why the missing context is immaterial.
[SCIENTIST]: Respond to the Politician's narrative. Point out the key omissions he is ignoring.
[JUDGE]: Once more — CONTINUE or STOP?
---
FINAL VERDICT [JUDGE]:
[REASON]: A brief explanation drawing on the arguments of both sides.
[VERDICT]: TRUE / HALF-TRUE / FALSE
Result: The model will show 2-3 rounds of debate. The Politician will build a narrative about traction, virality, and market interest. The Scientist will raise questions: how many of the 50k users are active? What's the cost per signup on a $40,000 budget? Do people come back? In the end, the Judge will most likely rule "HALF-TRUE," explaining that the metric is technically correct but hides the key indicators of growth quality.
Why It Works
The problem with the standard "find the weak spots." When you ask a model to critique something, it critiques what's written. But a half-truth is powerful precisely because what's written is true. The model sees a fact, accepts it as correct, and moves on. Missing context is invisible from a one-sided view.
Shifting the task from arguing to gap-hunting. Classic pro/con debates force both agents to work with the explicit content: one defends, the other attacks the same facts. RADAR changes the Scientist's job: he doesn't argue facts with the Politician — he looks for what didn't make it into the narrative. That's a fundamentally different generation mode.
The Judge as a quality filter. The Judge doesn't just summarize — he checks whether the discussion has run its course. If the arguments stop adding new information, the debate ends early. This matters: without a stop, the model starts rephrasing the same points over and over, and quality drops.
Control knobs: - Rounds — 1 round for a quick check, 3 rounds for deep analysis - Roles — swap Politician/Scientist for Marketer/Data Analyst in business cases; Defense Attorney/Investigator for legal claims - Verdicts — expand the options: "HIGHLY MISLEADING / SOMEWHAT MISLEADING / HONEST" - Scientist's instructions — add "look especially for: missing time context, missing comparison baseline, cherry-picked metrics"
Prompt Template
Run a structured analysis of the claim through a debate between three roles.
**Claim to analyze:** {claim}
**Context:** {known facts or background — or write "context unknown"}
---
ROUND 1. OPENING
[POLITICIAN]: Build a persuasive narrative in support of the claim.
Rely only on what is explicitly stated in the claim and context.
[SCIENTIST]: Check the claim for missing context.
What isn't said here? What data is absent? What could the chosen
wording be hiding? Don't attack the facts — attack the gaps.
[JUDGE]: Are the arguments sufficient for a verdict?
Answer: CONTINUE or STOP.
---
ROUND 2. REBUTTALS (only if the JUDGE said CONTINUE)
[POLITICIAN]: Respond to the Scientist's critique. Explain why
the missing context is immaterial to evaluating the claim.
[SCIENTIST]: Respond to the Politician's narrative. Point out the key
omissions he sidestepped.
[JUDGE]: CONTINUE or STOP?
---
FINAL VERDICT [JUDGE]:
[REASON]: A brief explanation drawing on the arguments of both sides.
[VERDICT]: TRUE / HALF-TRUE / FALSE
Placeholders:
- {claim} — a specific statement: a quote, a line from a pitch, a news headline, someone's announcement
- {context} — everything you know about the source, the numbers, the conditions. You can write "context unknown" — the Scientist will point out what data is missing on his own
🚀 Quick start — paste into the chat:
Here's the RADAR template for analyzing claims for hidden omissions.
Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask: which claim to analyze and whether there's any context — because without a specific claim, the roles can't start debating from different positions. It will take the structure from the template and adapt it to your task.
Limitations
⚠️ Subjective judgments: For claims with no objective truth ("this brand is interesting"), the method works worse — the Scientist has nothing factual to look for. It works on verifiable statements.
⚠️ Input quality: If the context is completely empty, the Scientist will list hypothetical gaps rather than real ones. The more context you give, the sharper the debate.
⚠️ The method doesn't replace fact-finding: RADAR analyzes what's already in the request. It doesn't go online or check the numbers itself. It's a tool for reasoning over data, not a verifier.
⚠️ A standard pro/con setup is worse for this task: If you swap the Politician/Scientist for an Advocate/Critic, you get a classic debate about explicit content. The omission-detection effect drops sharply.
Resources
Paper: Debating the Unspoken: Role-Anchored Multi-Agent Reasoning for Half-Truth Detection
Dataset/benchmark: POLITIFACT-HIDDEN (~15k political claims), AVERITEC
Code: github.com/tangyixuan/RADAR
Authors: Yixuan Tang, Yirui Zhang, Hang Feng, Anthony K.H. Tung — National University of Singapore / Shanghai Jiao Tong University