How Debates Work

Factual questions have objectively verifiable answers grounded in data, research, or established scientific or historical knowledge. The debate is looking for convergence on what is true. Example: "What are the proven health effects of caffeine?" Factual questions benefit most from models with strong knowledge recall and accuracy. Disagreements in factual debates often resolve quickly once models share their evidence.

Normative questions ask about ethics, values, policy, or what ought to be done. They do not have a single objectively correct answer. Instead, the debate is seeking alignment on a reasoned ethical framework and the conclusions it supports. Example: "Should AI-generated content be labeled?" Normative questions tend to produce the most substantive deliberation because different models may begin with genuinely different moral frameworks and must work through those differences explicitly.

Business questions are strategic or operational in nature. They concern risk, market dynamics, competitive positioning, or organizational decisions. Example: "Should a Series A startup expand internationally in year one?" The debate is seeking agreement on the key risks, trade-offs, and a recommended course of action. Business questions often benefit from a mix of analytical and creative models because the best answers combine rigorous risk assessment with lateral thinking about second-order effects.

Prediction questions are forward-looking. They ask about future outcomes, probabilities, or likelihoods. Example: "Will electric vehicles account for more than 50% of new car sales by 2035?" Consensus for prediction questions is measured by whether the models' confidence intervals and probability estimates converge. A debate where all models land within a similar probability range has reached consensus even if the exact numbers differ.

Brainstorm questions are open-ended. They seek a breadth of ideas rather than a single correct answer. Example: "What are unconventional marketing strategies for a B2B SaaS product?" Brainstorm debates are structured differently from analytical ones. Consensus is not about agreement on a single answer but about whether the space of ideas has been sufficiently explored. The debate ends when models stop introducing distinct new ideas that others have not already acknowledged.

Belief questions sit at the intersection of personal, cultural, or philosophical worldviews. They are not purely factual or normative but touch on foundational commitments about the nature of reality, consciousness, or human experience. Example: "Does consciousness require a biological substrate?" These questions rarely reach full consensus. Instead, the debate seeks to identify shared underlying principles or frameworks even when surface-level conclusions differ, and the synthesis reflects the genuine state of that partial agreement.

The adjudicator evaluates whether the models agree on the core facts, their relative importance, and the conclusions those facts support. Minor differences in emphasis or phrasing do not prevent termination. What matters is that no model is asserting a factually contradictory claim that others have not acknowledged and addressed.

The adjudicator evaluates whether the models have converged on a shared ethical framework and whether that framework leads them to the same conclusion. It is possible for two models to apply different ethical frameworks and still reach the same conclusion, which also satisfies this trigger. The key signal is that the models are no longer in substantive disagreement about what ought to be done.

The adjudicator evaluates whether the models agree on the key risks and their relative likelihood, and whether they are recommending the same course of action or a sufficiently similar one. Debates on complex business questions often converge on the risks quickly but diverge on the recommended response. Risk Stability is only satisfied when both the diagnosis and the prescription are aligned.

The adjudicator evaluates whether the models' stated confidence intervals and probability estimates overlap within an acceptable range. A model that says "60 to 70% likely" and a model that says "55 to 65% likely" have overlapping intervals and are converging. A model that says "30% likely" while others say "70% likely" has not converged and deliberation continues.

The adjudicator evaluates whether the debate is still generating distinct new ideas. In early rounds, models surface different angles and approaches. As rounds continue, new responses start to acknowledge and build on existing ideas rather than introducing genuinely new ones. Coverage Saturation is reached when the marginal idea contribution across models drops below a meaningful threshold, indicating the idea space has been sufficiently explored.

The adjudicator evaluates whether the models have identified a set of underlying principles or assumptions they share, even if their surface-level framings differ. Belief debates rarely produce full convergence on the conclusion. Framework Consensus acknowledges this and focuses instead on whether the models have articulated a shared intellectual foundation from which their different conclusions follow. This is the most nuanced termination trigger and the hardest to satisfy.

Debaters are the primary reasoning models. They receive the question, produce structured answers, defend their positions across rounds, consider counterarguments from other debaters, and revise their views when presented with compelling evidence or reasoning. The quality of the debate depends heavily on the diversity of the debater panel. Models from different providers, trained on different data with different optimization objectives, will surface genuinely different perspectives.

You can configure between 2 and 10 debaters depending on your plan. With 2 debaters you get a direct exchange between two perspectives. With 3 or more you get a richer discussion but more rounds are typically needed before consensus. For most questions, 3 debaters provides a strong balance between diversity and efficiency. 5 or more debaters is most valuable for brainstorm questions and complex strategic decisions where covering a wide surface area matters.

The adjudicator does not participate in the debate itself. Its role is to evaluate the state of consensus after each round and to score debater accuracy at the end. On the Free plan, consensus evaluation uses an algorithmic approach based on the structured fields in each debater's response. On Pro and Enterprise plans, you can configure any supported model as the adjudicator. A model-based adjudicator can assess nuances that are difficult to capture algorithmically, such as whether two models are genuinely aligned on a normative framework or are simply using compatible language to describe incompatible positions.

Choosing a strong adjudicator matters most for normative, belief, and complex business questions where consensus is qualitative rather than quantitative. For factual and prediction questions, the algorithmic approach on Free is often sufficient because convergence is easier to measure.

Once the debate reaches consensus or exhausts its round limit, the synthesizer writes the final answer. It receives the full debate transcript, including the blind round responses and all deliberation round responses, presented anonymously. Its task is to distill what the debaters agreed on into a single clear answer that does not privilege any individual model's phrasing or framing.

The synthesizer is always a different model from the debaters. This separation ensures that the final answer is a genuine distillation rather than a reproduction of whatever the most influential debater said. The synthesis is not a summary of the debate process. It is a direct answer to the original question that reflects the consensus position.

Stance Alignment measures how closely the debaters' final positions agree at the level of their stated conclusion. Two models that both conclude "yes, this policy should be adopted" have high stance alignment. Two models where one says "yes with significant caveats" and another says "no under current conditions" have low stance alignment. This is the most direct measure of whether the models are actually agreeing.

Empirical Overlap measures how much the factual claims and cited evidence from each model intersect with those of the others. Models that are citing different studies, different data points, or different historical examples to support their answers may reach the same conclusion but are doing so on different evidential foundations. High empirical overlap means the models are working from a shared factual basis, which makes the consensus more robust.

Framework Agreement measures whether the models share a common analytical or reasoning framework. This dimension matters most for normative, business, and belief questions. Two models can agree on a conclusion while applying entirely different reasoning frameworks, and the synthesis of such a debate needs to acknowledge this. Conversely, two models that apply the same framework are likely to resolve remaining disagreements in subsequent rounds because they are operating from the same analytical foundation.

Confidence Convergence measures whether the models' stated confidence scores are approaching one another across rounds. A panel where one model reports 0.95 confidence and another reports 0.45 confidence is not converging on a reliable answer, even if their stated positions are similar. Genuine consensus requires not just agreement on what is true but approximate agreement on how certain that truth is.

Factual Accuracy measures how well the model's claims aligned with the consensus and with verifiable facts. A model that asserted claims that were later corrected by other debaters, or that the adjudicator identifies as factually contested, will score lower on this dimension. A model whose claims were consistently validated and built upon by the debate scores higher. This dimension is most meaningful for factual and prediction question types.

Logical Consistency measures whether the model's arguments were internally coherent and free of contradictions across rounds. A model that reversed a position without acknowledging the reversal, or that held contradictory claims simultaneously, will score lower. This dimension rewards intellectual honesty: a model that explicitly identifies when and why it changed its mind scores higher than one that silently shifts position without acknowledging the change.

Evidence Quality measures how well the model supported its claims with reasoning, examples, analogies, or data. A model that asserts conclusions without supporting them scores lower than one that explains its reasoning in depth. This dimension does not require the model to cite external sources. Strong logical argumentation, well-constructed analogies, and clearly stated inference chains all contribute positively to Evidence Quality.

The Trajectory Score is distinct from the three accuracy dimensions. It measures how constructively the model engaged with the debate process itself across rounds. A model that started with an incomplete initial answer but meaningfully revised it in response to other debaters' arguments scores higher on trajectory than one that had a strong initial answer but ignored everything the other models said. A high Trajectory Score indicates a model that is genuinely responsive to argument rather than one that simply outputs its initial conclusion and waits for the debate to end.

Trajectory Score matters because the value of a debate system comes precisely from models updating on each other's reasoning. A model with a high Trajectory Score is contributing to the process. A model with a low Trajectory Score, even if its initial answer was correct, is not making the debate better.

In Auto mode, DebateTalk classifies your question and then selects the best available models for that question type. The selection logic applies two priorities simultaneously. First, it favors models that have historically performed well on questions classified in the same domain, drawing on accuracy scores accumulated across all debates on the platform. Second, it enforces diversity across AI providers. A debate panel will not be filled with models from the same provider, because models from the same provider tend to have similar training backgrounds and will naturally agree more quickly, reducing the value of the debate.

Auto mode is the recommended starting point for most use cases. It removes the need to research individual model capabilities and regularly incorporates new models as they become available and establish a performance track record.

In Manual mode, you configure the debaters, adjudicator, and synthesizer yourself via the model configuration parameters in the API request. You can mix any combination of supported models. Manual mode is appropriate when you want to compare specific models head-to-head, when you are building a product that always uses the same fixed panel, or when you are benchmarking a new model against established ones to understand how it performs relative to the field.

Manual mode requires more configuration knowledge but gives you complete control over the composition of the panel. You are responsible for ensuring diversity: if you configure three models from the same provider, the debate may converge faster but the consensus will reflect a narrower range of perspectives.

running is the active state. The debate is currently processing and streaming SSE events to the client. Events arrive in order: round events as each debater completes its response, consensus events after each round, a synthesis event when the synthesizer finishes, and an accuracy event with the final scores.

completed means the debate finished normally. Either the consensus threshold was crossed and synthesis ran, or the round limit was reached and synthesis ran anyway. The history record is fully populated and all SSE events were emitted. This is the expected terminal state for a successful debate.

failed means an unrecoverable error occurred during the debate. The most common cause is all debaters in a given round returning errors simultaneously, for example when multiple model provider APIs are unavailable. When this happens, the debate transitions to failed immediately rather than attempting to continue with partial responses. A partial transcript may be stored in the history record but synthesis does not run.

aborted means the client disconnected before the debate finished. SSE connections are stateful. If the connection drops mid-debate, the backend records the state at the point of disconnection. A partially completed debate in aborted state will have whatever rounds completed before the disconnect stored in history, but synthesis will not run.

Authenticated debates are stored in history regardless of their final state. A failed debate and a completed debate are both stored. The status field in the history record tells you the terminal state. Unauthenticated requests are not stored. The debate history API returns the full transcript, consensus scores per round, synthesis, accuracy scores, and metadata including the question classification, the number of rounds run, and the terminal state.

The last SSE event in every debate is a final event. It contains the debate ID, the terminal status, a reference to the synthesis, the overall consensus score at termination, and a brief summary of the debate run including how many rounds were completed and whether consensus was reached or the round limit was hit. The final event is your signal that the stream is complete and the history record is fully written.