People bring AI questions they intend to do something with. A proposal to send. An analysis to rely on. A decision that becomes expensive to reverse. They need an answer, but they also need someone to notice when the question or the first response rests on a weak assumption.
That is the job Pingpong is designed to do. It runs a selected sequence of models and returns one final response, with the earlier passes available to inspect.
The handoff is part of the method
The first model responds to the question. Each later model receives the original question and prior responses, together with a review frame that permits agreement, correction, restructuring, or a different approach. An earlier answer is material to assess, not an instruction to endorse.
This framing is applied again at each later pass. A review should be able to challenge the premise, keep a sound conclusion, or decline to support a claim that lacks evidence. Our thesis is that the conditions of review affect how well a model's capabilities are used.
GrokFirst response
PerplexityReview
ChatGPTReview
GeminiReview
ClaudeReview
Order and model selection are configurable. Later passes receive the review frame; the first does not.
Independent judgment, with shared context
Independent does not mean isolated. Pingpong's reviewers see earlier answers. That enables direct correction, but it also creates a risk of anchoring. We do not describe the sequence as blind review or as statistically independent votes.
Chain-of-Verification shows why controlling exposure to a draft matters. Mixture-of-Agents shows a different use of shared context: previous outputs become inputs to further generation. Pingpong is a sequential review system, not an implementation of either paper.
Candor is not compulsory disagreement
The sycophancy literature gives a reason to question agreeable answers. But a model that automatically objects is no better at listening to evidence. Recent work on rational updating makes the distinction explicit: resist unsupported pressure while remaining willing to correct an error.
We use review instructions at inference time. We are not claiming to have retrained the underlying models, made them unbiased, or eliminated sycophancy. The intended behavior is considered judgment, including agreement when the reasoning supports it.
The framework has to earn its place
Adding models adds latency and cost. Research on repeated use of a strong model and debate versus voting gives us alternatives that a critique chain should be tested against. Provider diversity and a longer transcript are not sufficient evidence of improvement.
A convincing evaluation would compare the best single model, repeated sampling, an unframed chain, and the framed chain at matched budgets. It would count mistakes introduced as well as mistakes corrected. It would test whether important objections survive into the final response.
One usable answer, without hiding uncertainty
A person or an agent needs an output it can move forward with. Sometimes that is a revised recommendation. Sometimes it is a missing fact that prevents a responsible recommendation. A clean answer should organize the uncertainty, not erase it.
That is what we mean by decision insurance: review before commitment. It is a description of the job, not a financial insurance product or a guarantee. Consequential actions still need evidence, appropriate expertise, and explicit authorization.