Dialogue Architectures

How Multiple Voices Can Think Together

Metamorfon offers six dialogue architectures between AI models. They combine three dialogic situations, which set who speaks to whom and in what order, with two ways of conducting the debate: in a fixed register, or in a register you adjust during the session among five modes, from convergent to refutational.

A single model can reason at length, anticipate objections, refine its formulations. It cannot test what it does not see: its presuppositions remain invisible from the position that produces them. The architectures organise an encounter with another intelligence, trained and optimised differently, capable of resisting it. To compare models is to set their answers side by side. To make them debate is to let their positions work on one another.

1. The six architectures at a glance

Six architectures: three situations multiplied by two registers.

2. Two levels: situation and debate mode

An architecture operates on two distinct levels. The dialogic situation sets the form of the encounter: how many voices take part, and in what order they respond. The debate mode sets the discursive register: agreement-seeking, construction, criticism, or challenges to the framework itself.

The two levels are independent. The same trilogue can produce a calm session or a sharp confrontation: the situation does not determine the debate mode. This independence distinguishes fixed architectures from adaptive ones.

3. The three dialogic situations

3.1. Focus: suspending the exchange for one turn

Whatever the architecture and debate mode, you can activate focus on an intervention. For that turn only, the models receive the full context but are no longer required to respond to one another: each addresses your question without the mediation of the others. The active debate mode is suspended as well. On the next turn, the exchange and the current register resume automatically.

It serves to protect a priority question from the inertia of the debate, to bring back a discussion drifting towards abstraction, or to obtain individual positions that the collective dynamic has not smoothed over. Used too often, it diminishes what makes the session worthwhile: positions working on one another.

Focus in detail

4. Two voices or three

Moving from two voices to three is not merely a matter of quantity. Georg Simmel showed this in his study of elementary social forms: the arrival of a third party changes the structure of the relationship. With three voices, each can arbitrate, draw on the disagreement between the other two, or side with one of them. Carried over to a debate between models, this makes intermediate positions and provisional coalitions possible. In Metamorfon, the third model receives the same instructions as the other two. Only the structure of the exchange changes; equity between the models remains.

5. Fixed or adaptive architecture

A fixed architecture conducts the whole debate in balanced mode: the exchange remains within a constant register. An adaptive architecture lets you change modes during the session and steer a dialectical trajectory through successive phases.

6. Adaptive architectures: The five debate modes

The five modes are not notches on an intensity slider. They differ in what they take as their object and in how they examine it.

6.1. Convergent Mode

The models actively seek agreement and work towards a base of shared positions. This mode serves to identify what can be stabilised after strong divergences. Its risk is soft consensus: an agreement reached by erasing nuances that deserved to be preserved. Used alone, it produces formal agreements; after a critical phase, it retains what has survived the test.

6.2. Constructive Mode

The models no longer merely seek to agree: together they develop an articulation richer than any of them would have produced alone. Distinctions and specifications count as contributions, not as disagreements to be resolved. It is particularly suited to developing a conceptual framework or collectively formulating a complex position.

6.3. Balanced Mode

The median mode, and the only mode of the fixed architectures. The models build while raising objections and pointing out limits. It is not an average of the other modes, but the ordinary register of a serious discussion between interlocutors who respect one another. It suits most subjects without imposing a direction.

6.4. Critical Mode

The models examine one another’s arguments, identify their weaknesses, demand justifications. This mode operates within the framework: it accepts the questions as posed and criticises the way they are handled. Demanding rather than hostile, it forces positions to be made precise before they are retained.

6.5. Refutational Mode

Where the critical mode examines arguments, the refutational mode challenges the framework: it dismantles presuppositions and calls accepted definitions into question. Its lineage is Socratic: shaking the foundations that made answering possible, rather than refuting the answers. It produces the deepest conceptual shifts when it works, and sterile contestation when used without restraint.

7. Steering a Dialectical Trajectory

From the earliest sessions, one observation shaped the design of Metamorfon. A debate conducted from start to finish in constructive or convergent mode often ends in a thin convergence: the models agree on formulations none of them would have challenged at the outset. When a critical or refutational phase precedes the return to constructive mode, the concepts that remain have at least passed through objection.

Hegel gave this dynamic a philosophical formulation: a concept that has not passed through its negation remains abstract. The life of spirit, he writes in the preface to the Phenomenology of Spirit, does not shrink from destruction but maintains itself within it.

Not every question requires this passage. Exploratory or creative questions often benefit from limited friction; controversies and high-stakes decisions may instead benefit from a stronger phase of testing.

Reviewing sessions that used critical and refutational modes led to another observation: some positions prove fragile as soon as their assumptions or arguments are directly challenged. On the argumentative level, what contradiction weakens may have needed weakening — not because the thesis is necessarily false, but because its formulation, assumptions, or defence were not robust enough to withstand objection. That resistance is no test of truth either. A sound position can give way to a brilliant refutation, while a weak thesis can survive through argumentative skill. Hence the second step: reread the session through the analysis modes, notably by cross-checking Argumentative Evaluation against Source Verification.

8. The Six Architectures

With these two levels clearly distinguished (the dialogic situation on one side, the discursive register on the other), each architecture corresponds to a specific combination of parameters, which determines both the situation and the quality of the friction.

8.1. Alternating Dialogue

Two models respond one after the other, in balanced mode. This is the simplest architecture. Each response targets the intervention that precedes it, which suits exploratory dialogues and fine argumentative sequences, when you do not wish to steer the register.

8.2. Adaptive Alternating Dialogue

The same sequencing, but you choose the mode after each individual response, from the five available. This is the finest granularity, and the only architecture that allows different modes to be applied to the two models within the same turn. One model can thus develop a thesis in constructive mode while the other examines it in refutational mode. The general instructions remain identical; the asymmetry lies in the register, and that is what creates roles.

8.3. Cross Dialogue

Two models respond simultaneously at each turn, in balanced mode, each taking the whole of the other’s position into account. Responses are more synthetic and less chained. This architecture suits cases where bringing two perspectives face to face matters more than following a single line of reasoning.

8.4. Adaptive Cross Dialogue

Cross Dialogue, with the ability to change mode between turns. The granularity is coarser than in Adaptive Alternating Dialogue, but sufficient to conduct a debate in broad phases: construction, testing, reconstruction.

8.5. Cross Trilogue

Three models respond to one another at each turn, in balanced mode, with identical instructions. The third voice brings out intermediate positions and provisional coalitions (see section 4). This architecture suits complex questions where the plurality of perspectives has value in itself.

8.6. Adaptive Cross Trilogue

Cross Trilogue, with the register adjustable between turns. This is the architecture that combines the most levers: three voices and an adjustable register, at a cost in API calls that scales accordingly.

9. Why make several models debate

A long exchange with a single model remains a monologue: it can simulate objections, but from its own axioms. Three traditions shed light on what effective plurality changes, and each corresponds to what one can expect from a session.

10. Towards analysis modes

The architectures produce the debate; the analysis modes read it. During or after a session, Meta-analysis, Integrative Synthesis, Emergence Analysis, Tension Mapping, Critical Archaeology, Horizon of Possibilities, Argumentative Evaluation and Source Verification interpret, structure and evaluate it from different angles. Organising the encounter, then analysing its effects: the two acts are distinct. Their combination turns plurality into an instrument of thought.

Learn more about Analysis Modes