Not an everyday AI assistant. ChatGPT, Claude, Grok or Gemini work in the register of request and response: a single interlocutor, adjusting to the user as the exchange goes on. Metamorfon brings several distinct models together and shifts the relation: they no longer answer only the user, they answer one another. From one turn to the next, their positions reinforce each other, shift, contradict each other, recompose.

Not a research assistant. A research assistant has one function: to find information and report it. Metamorfon calls on web search when a debate depends on recent facts, but there information remains a material. Everything happens after it arrives: how one model interprets it, how another contests it.

Not a productivity tool. Productivity tools speed up tasks one already knows how to do. Metamorfon slows down on purpose. Formulating the question, letting the debate unfold, isolating a point, changing debate mode, rereading the analyses: this time is no side cost. The object of reflection changes while it passes.

Not a benchmark. A benchmark measures performance on tasks fixed in advance. A sustained dialogue exposes a conduct: how a model defends a position, answers an objection, holds or revises its commitments, and lets its presuppositions show when another attacks its reasoning.

Not a content-generation tool. A session first produces a trajectory: initial positions, objections, shifts, dead ends, corrections, sometimes a concept taking shape. The analysis modes then reread that trajectory from several angles.

These five gaps outline its category: Metamorfon is an instrument of thought.

The phrase invites an objection: why would making several models talk to one another improve anything at all?

Nothing guarantees it. Hence the method.

Contradiction probes; it does not prove

A large language model excels at producing a coherent answer from a context. That is where its usefulness comes from. So does the difficulty, once it is used for thinking.

A fluent, well-structured, apparently documented answer can rest on a fragile premise. The model skilfully unfolds the consequences of a questionable framing, because nothing in the exchange forces it to go back to it.

The training of assistants accentuates the trait. They are not only taught to produce plausible language: their post-training incorporates human preferences about what counts as a useful answer. Work on sycophancy (Sharma et al., ICLR 2024) shows that, under certain conditions, this optimization pushes a model to side with its user’s beliefs rather than hold to the accurate answer. The phenomenon is neither uniform nor inevitable, and models are also trained to resist it. It is real enough for the major providers to devote specific evaluations to it.

Asking the same model to criticize itself settles nothing by default. Without reliable external feedback, self-correction often fails to improve a line of reasoning and sometimes degrades it (Huang et al., ICLR 2024). A second answer from the same system is not a second pair of eyes.

Hence the research interest in debate between models. When several instances confront their answers and then revise them, reasoning and factuality improve on several tasks (Du et al., ICML 2024): the objection brings the model information absent from its own trajectory and sets it a new problem.

The gain is not automatic: debate systems do not reliably outperform methods without debate, such as a vote among several answers from the same model, and prove more sensitive to how their parameters are set (Smit et al., ICML 2024). In a controlled study on a logic puzzle, success depends first on the agents’ reasoning strength and on the diversity of the group, and majority pressure suppresses independent correction (Wu et al., 2025 preprint). Debate does not guarantee correction.

Nor are the models independent witnesses. Their corpora overlap, as do their cultural references, and certain presuppositions run through all the texts they learned from. Measured across more than 350 LLMs, their errors are substantially correlated, and those of the best-performing models remain so across distinct providers and architectures (Kim et al., ICML 2025). Three models repeating the same figure do not make three sources.

Plurality, then, is not proof. It creates occasions for testing. A claim passes into the hands of an interlocutor who is a stranger to the chain of arguments that produced it; an objection forces a premise left implicit to be spelled out. Two positions believed compatible turn out to contradict each other as soon as they share the same space of discussion.

Metamorfon does not make AI answers true; it arranges conditions in which error has more chances of being noticed.

A methodological framework

Making several LLMs debate is still only a technical setup. The instrument begins with decisions: how the models meet, what they receive, what is expected of them, how their answers are examined, where the user steps in. In other words, with a method.

Metamorfon’s method operates at four levels: the instructions given to the models, the architecture of their exchanges, the debate mode applied to a turn, the analyses that go back over the session. At none of these levels does it dictate the right answer. It sets the conditions of the examination.

The concrete workings of this orchestration — dialogue architectures, debate modes, Focus, and analysis — are developed in Make Two AIs Debate: Orchestrating and Analyzing the Dialogue. Here, the emphasis is less on how to conduct a session than on what the framework does: what it allows us to put to the test, make visible and, at times, recompose.

The method rests first on the instructions. Asking an AI to “reveal the biases” in a text assumes it contains some; asking it to “show why a thesis is false” fixes the destination before the inquiry. The instruction then manufactures part of the result it claims to discover.

The spirit of the instructions comes down to one rule: prescribe a form, never a conclusion. Axes, an order of examination, a discipline of restitution. Their wording removes, one by one, the words that incline interpretation: we call this “lexical depopulation.”

A critical method therefore does not consist in ordering models to be critical. It puts propositions to the test without having decided in advance which one will have to give way.

The dialogue architectures form the second level of the method. They set who speaks to whom, in what order, and what each model receives from the others before answering.

The debate modes then change the rules of the game. Convergent, constructive, balanced, critical, refutational: not five notches on an intensity slider, but five distinct operations on the same material. The critical mode examines arguments within the accepted framework; the refutational mode challenges the framework itself.

In adaptive architectures, the mode therefore varies while the question stays the same. A proposition built during one turn of dialogue can be handed over, on the next, to a search for its weak points. Conversely, a highly critical phase can give way to a constructive mode, in which the models rebuild from what has held.

The method does not seek disagreement for its own sake. It aims to make disagreement, agreement and revision informative.

An argumentative discipline

Organizing turns of speech is not enough. A debate can be spectacular and mediocre. Two interlocutors contradict each other without ever answering the other’s argument. A model shifts a definition along the way, changes its standard depending on whether it is judging its own thesis or another’s, makes its position gradually irrefutable. Contradiction no more produces reason than plurality produces truth.

Argumentative Evaluation bears on this point. It does not say who is right on the substance. It observes the conduct of the reasoning: did the objection receive an answer? Does the conclusion follow from the premises advanced? Is a justified correction acknowledged? Does a model demand from its interlocutor a proof it does not itself provide? Call this an argumentative discipline: consistent criteria, fair treatment of the opposing argument, capacity for revision, the distinction between objection and refutation, refusal to shift the terms of the debate in silence.

These rules defend no doctrine; they make the defense of every doctrine more demanding. The stakes are sharper with language models, much of whose power lies in the capacity to produce fluent and convincing discourse. Rhetorical force and logical soundness are two properties. An instrument of thought must be able to tell them apart.

Another model is not another source

Argumentative examination leaves another fragility intact: a well-built line of reasoning can rest on a false premise. A misreported figure, a non-existent study, a conclusion wrongly attributed to a source settle easily into a coherent argument. Debate does not guarantee their detection: systems trained in a partly shared informational environment produce correlated errors.

The methodological framework must therefore, at some point, step outside the dialogue. Source Verification does not check every sentence. It isolates the verifiable claims that carry the demonstration (decisive figures, studies invoked as evidence, attributions to an author, works, dates) and checks them against external sources.

An argument is audited from within, through its coherence, whereas a factual claim requires a confrontation with what lies outside the dialogue.

An accurate source does not repair bad reasoning; rigorous reasoning does not turn a false datum into a fact.

Argumentative Evaluation and Source Verification answer two distinct vulnerabilities. Neither is a truth machine. Together, they multiply the points at which errors (faults of reasoning, invented or over-interpreted sources) become visible.

The user is part of the problem

A deeper difficulty remains: one can examine the answers, the arguments and the sources without ever questioning the question that produced them all. Yet every question already carries a representation of the problem.

Asking “what does a society lose when…” assumes a loss. Asking what solution to bring to a phenomenon already makes it a problem. Choosing a legal vocabulary rather than an economic one, an individual rather than a systemic one, marks out the space in which the models will look for their answers. The flaw is unavoidable: all thinking starts somewhere. That starting point, however, can be examined.

Critical Archaeology traces back to the presuppositions lodged in the initial question and in the user’s interventions: concepts already installed, oppositions taken for granted, a vocabulary that lights up certain objects and leaves others in shadow. It does not grade the question. It brings to light the decisions the question had made before the first answer.

The usual distribution of roles is altered. In an ordinary chat, the user asks and the model is evaluated. Here, the one who asks enters the field of examination in turn. The critical method becomes reflexive.

Errors that only duration reveals

An isolated answer shows what a model produces in a given situation. It does not say what becomes of that position when it meets an objection, a contradictory datum or a reformulation of the problem. Producing an answer and sustaining it in the face of objection are two distinct tests. Some fragilities are therefore interactional: they exist, or become visible, only over the duration of the exchange.

A model inflects its position by small touches and, a few turns later, defends something other than its initial thesis. The shift may be reasonable and never acknowledged.

It would be wrong to conclude from these “silent revisions” that an LLM cannot reread itself. They echo a difficulty documented by the self-correction research cited above: a model is not, for itself, a sufficiently independent critic.

The duration of the dialogue then acquires a value of its own. The question is no longer only “is this answer good?” It becomes: is this position still compatible with the one the model held four turns earlier? Which argument made it change? Did it change its conclusion without touching the reasons that supported it?

Tension Mapping identifies the disagreements that hold. Meta-Analysis looks for the structure beneath the positions: shared axioms, diverging conceptions of proof, common blind spots beneath the apparent disagreements.

A noisy controversy sometimes turns out to be very homogeneous. Three models fight over the answers and silently accept the same question.

The method names no winner

The phrase “AI debate” calls up the image of two positions and a winner. The image covers only a special case. A persistent disagreement sometimes teaches more than a victory. A convergence gives way at the first refutation, or withstands several. An objection does not always destroy a thesis: it forces it to specify its domain of validity.

The method seeks less a “winner” than a change in the state of the problem. Here the instrument’s second function appears: testing an idea sometimes leads to transforming it. That is where the application’s name comes from.

From stress test to laboratory

Large language models are remarkably good at recombining the elements of their context. A distinction, an objection, an analogy, one more constraint, and the next answer reorganizes the whole. A one-on-one chat already shows this.

A multi-model exchange adds a source of variation: part of the new context comes from another system, engaged in its own argumentative trajectory.

Model A proposes a distinction. Model B accepts only part of it and introduces an unforeseen objection. A, or C, is no longer answering only the initial question: it is answering this new object. A few turns later, the initial opposition has sometimes disappeared in favor of a distinction no one had formulated in that way at the outset.

This has been called a “third thought.” The expression should say no more than what can be observed: nothing shows that no model would have produced the result alone. Its genealogy within the session, however, can be established: the final proposition emerges from transformations of which the exchange keeps the trace.

Hence the laboratory metaphor. A session is not a controlled experiment; one can, however, vary the conditions of the interaction and observe what happens to ideas. The architectures change the circulation of information, the debate modes the argumentative rule. An analysis spots a blind spot and draws a new question from it. The user steps in when the collective dynamic itself becomes a constraint.

Isolating a question without erasing the debate

Focus mode condenses this logic. It does not add a sixth mode to the other five; it operates on another plane. For one turn, the models keep the full context of the session but are no longer required to respond to one another: each addresses the question asked without the mediation of the others. The active debate mode is suspended as well, then resumes on the next turn.

This suspension isolates a variable. A collective discussion has its inertia. The further it goes, the more certain concepts dominate; an objection is absorbed by the common framework before it has been examined for its own sake. The models drift toward abstraction until they lose sight of the question the user was trying to resolve. The debate also grows denser faster than a human reader can follow.

Focus introduces a local break without erasing the history that precedes it. Its value lies in its rarity: activated at every turn, it would remove the property it protects, positions working on one another. In the laboratory, it serves as an isolation chamber.

Tracing an emergence

Emergence Analysis looks for the transformations born of this dynamic. The word calls for a strict definition. It does not designate a concept unprecedented in the history of ideas, nor one that no model would have produced in another conversation. It designates a concept, a distinction or a framework absent from the starting point, whose formation can be traced through the successive contributions of the session.

This genealogy matters more than a claim to originality. In it one reads what one model brought, what another contested, the leads abandoned, the transformation that broke the deadlock.

Integrative Synthesis performs the complementary operation: from the plurality of contributions, it reconstructs a coherent whole. One looks at the transformation, the other at the structure that can be built on its results.

An answer agreed, then withdrawn

A session on what extreme wealth buys shows the passage from critique to recomposition.

The question offered two branches: does extreme wealth buy more of what ordinary wealth buys, or something different in kind? Answering without seeing one another, the three models converge at once: different in kind, and the difference is power. That unanimity could have stood as the answer. The rest of the session turns it into a problem.

The user brings in a single figure: the owner of a company town. What the debate had attributed to billionaires already existed at that scale. Leaving, writing the rules, keeping issues off the agenda, deciding alone: each power is then examined in turn, and none comes out intact. One model drops the phrase it had used from the start; the one that had opened on “a different kind of good” ends on “not a kind apart.”

This retraction does not leave a void. What remains is narrower and firmer: scale relative to the arena, speed, reach. The billionaire emerges as a company-town owner at the scale of a country, where courts, a press and rival fortunes still stand in the way.

The session does not deliver a better answer to the initial question. It delivers a better question: no longer whether extreme wealth differs in kind, but how long a power can last without having to be reauthorised. The relation between Metamorfon’s two dimensions lies there: critical examination tests, the laboratory recomposes.

Analysis opens the next experiment

The analysis modes do not necessarily mark the end of a session. A Meta-Analysis reveals an axiom no one had questioned. Critical Archaeology shows that a possibility was excluded by the initial formulation. Emergence Analysis points to the new distinction on which to focus what follows. Horizon of Possibilities lists the directions the debate made accessible without entering them. Each of these analyses ends with a question: the one the analyzing model would put to the participants at this stage of the debate.

That question can be fed back into the session as it stands, or reformulated by the user. When the new question risks being absorbed by the inertia of the dialogue, focus mode isolates it for one turn; the collective exchange then resumes, with, in its history, an element examined for its own sake.

The setup forms a loop: dialogue → analyze → reformulate → test again. Debate no longer serves only to produce an answer: it becomes an iterative inquiry.

The human remains responsible for the method

This loop does not run by itself. The user chooses the initial question, the models, their own interventions, the debate modes, the analysis modes. They decide that a disagreement deserves to be dug into, that a claim must be verified, that a discussion has locked itself into too narrow a frame. Above all, they decide what the result means.

No consensus among models makes a proposition true. No Argumentative Evaluation establishes on its own that a conclusion is right. No Source Verification removes the need to interpret the sources. No Emergence Analysis guarantees that what emerged is fruitful.

The methodological framework increases the authority of artificial intelligence less than the number of handholds offered to human judgment. It makes visible divergences, revisions, presuppositions, factual fragilities, new concepts, framing effects, questions still open.

This abundance makes reflection more demanding. It sometimes removes a certainty without supplying the one that will replace it. A critical method promises nothing else: it does not eliminate uncertainty; it prevents it from hiding behind the fluency of an answer.

From the answer to the trajectory

Conversational interfaces have accustomed us to seeking a better answer: better written, more precise, better documented, from a more powerful model. Metamorfon explores another variable: the framework in which the available intelligence is put to work.

A single model, several parallel answers, an adversarial debate, a collective construction, a question passed through several successive debate modes: so many different operations, even when they mobilize exactly the same LLMs.

Metamorfon does not promise a superior intelligence. It offers a methodological framework for working with imperfect intelligences.

The relevant unit is no longer only the question-answer pair. It becomes the trajectory: a proposition, an objection, a resistance, a verification, a revision, a shift of frame, sometimes a distinction absent at the start. Then a question that can no longer be asked in quite the same way.

Metamorfon is an instrument of thought in this sense.

An instrument for testing ideas. A laboratory for transforming them.

References

• Sharma, M., Tong, M., Korbak, T. et al., “Towards Understanding Sycophancy in Language Models”, ICLR 2024.
• Huang, J., Chen, X., Mishra, S. et al., “Large Language Models Cannot Self-Correct Reasoning Yet”, ICLR 2024.
• Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., Mordatch, I., “Improving Factuality and Reasoning in Language Models through Multiagent Debate”, ICML 2024.
• Smit, A. P., Grinsztajn, N., Duckworth, P., Barrett, T. D., Pretorius, A., “Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs”, ICML 2024.
• Wu, H., Li, Z., Li, L., “Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning”, arXiv preprint, November 2025.
• Kim, E., Garg, A., Peng, K., Garg, N., “Correlated Errors in Large Language Models”, ICML 2025.

L’article Metamorfon, an Instrument of Thought est apparu en premier sur Metamorfon.

You’ve probably seen the scene: two artificial intelligences wired to each other, left to exchange until the conversation drifts, repeats itself, or produces a jargon no one can follow anymore.

The demonstration establishes that a dialogue between AIs is possible. It does not tell us what such a dialogue is good for. Making two AIs debate becomes a method when the exchange is organized so that positions, objections, shifts, and disagreements can be observed — and then analyzed. Metamorfon builds this process around two gestures: orchestrating the dialogue, then analyzing its trajectory.

The question of why this confrontation can become an instrument of thought — and why it offers no guarantee of truth — is developed in Metamorfon, an Instrument of Thought. Here the question is more practical: how can several models be made to debate without their conversation becoming shapeless material?

Why make two AIs debate?

Query one model and you get an answer. Query several models separately and you get several answers.

Make them respond to one another and you get something else: a trajectory.

A claim meets an objection. The model that made it can maintain it, refine it, give up part of its reasoning, or shift the problem. Another model can introduce a distinction that was absent from the first turn. An initial convergence can strengthen or break apart once it is subjected to criticism. This temporal and interactional dimension is what distinguishes a debate from a simple juxtaposition of answers.

Research on multi-agent debate confirms that, under some protocols, this kind of confrontation can improve reasoning and factuality (Du et al., ICML 2024). It also shows that the benefit is not automatic: debate strategies do not consistently outperform simpler approaches such as ensembling or self-consistency, and their performance depends strongly on the protocol used (Smit et al., ICML 2024).

So multiplying voices is not enough. The form of their interaction matters.

A session on AI monoculture and the convergence of thought illustrated this in an almost reflexive way. Three models invited to discuss the risk that AI might homogenize ways of thinking progressively converged in their own ways of reasoning: similar references, related distinctions, a shared grammar of justification.

The point was not that this convergence was enough to establish the existence of an AI monoculture, but that it made some of its possible mechanisms observable.

The problem with improvised debate

Putting several models into conversation is not enough to produce a good debate.

The first risk is drift. As turns accumulate, models respond increasingly to what their interlocutors have just said and can gradually move away from the problem that motivated the exchange.

This is not merely an informal observation. Becker and colleagues studied it as problem drift in extended multi-agent debates: across several categories of tasks, the discussion can progressively move away from the original problem, especially when exchanges stop producing clear progress or feedback between agents becomes uninformative (Becker et al., EACL 2026).

A second risk concerns positions. If nothing requires a model to account for its previous commitments, it can gradually alter its thesis without indicating what led it to change. A justified revision is interesting; a variation without memory is much less so.

The third risk is a conversation that becomes difficult to read afterwards. A succession of texts can be intellectually rich while making it very difficult to reconstruct what actually happened: where an objection appeared, which distinction held, which argument triggered a revision.

Orchestrating the debate means making that trajectory observable.

Orchestrating: giving the exchange a form

The first decision concerns the dialogue architecture: who speaks to whom, in what order, and with what information.

In Metamorfon, two models can exchange response by response in an alternating dialogue, or respond simultaneously after reading one another’s contributions from the previous turn in a cross dialogue. A third model can be added in a cross trilogue. These architectures do not merely change the presentation. They change how information circulates through the debate.

Then come the instructions given to the models. Their purpose is not to determine what the models should conclude, but to set the rules of their participation. The Spirit of the Instructions rests on this distinction: prescribe a form, never a conclusion.

A model is therefore not assigned a thesis chosen in advance. Once it has taken a position, however, the continuity of the session makes it possible to examine what it does with that position.

Holding a position does not mean clinging to it. A position can evolve if an objection lands. What matters is that the shift remains intelligible: that we can identify what changed and, where possible, why.

Changing the regime without changing the question

Orchestration also concerns the regime of the dialogue. It can be fixed or adaptive.

In adaptive architectures, Metamorfon distinguishes five debate modes: Convergent, Constructive, Balanced, Critical, and Refutational.

These are not five increasing degrees of aggressiveness.

Convergent mode looks for what can be held in common. Constructive mode asks the models to build on one another’s contributions. Balanced mode keeps cooperation and objection in play at the same time. Critical mode examines arguments within the accepted framework. Refutational mode allows the framework itself to be questioned: presuppositions, categories, or premises the participants had so far taken for granted.

In an adaptive architecture, these regimes can change during the same session.

The question remains the same; the operation performed on it changes.

This steering is not automatic. The models apply the regime they are given; the choice of regime belongs to the user steering the session. The user observes what the previous turn produced and decides what now deserves to be built, challenged, or displaced.

Reframing without erasing what came before

Even a well-orchestrated prolonged debate acquires inertia.

By continually responding to arguments already present, the models can absorb a new user question into the framework they have collectively built rather than examining it on its own terms.

This is where Focus mode intervenes.

For one turn, the models retain the full context of the session but are no longer required to respond to one another. Each addresses the user’s intervention directly; the collective regime resumes on the following turn.

Focus is therefore not a sixth debate mode. It operates orthogonally to the debate modes: it temporarily suspends the interaction in order to protect a question from the inertia of that interaction.

Its usefulness exists precisely because drift exists. But it should remain occasional: used on every turn, it would remove what makes the multi-model setup valuable — positions working on one another.

Analyzing: reading what the debate produced

A completed session contains more than its conclusions.

It contains initial positions, objections, concessions, persistent disagreements, changes of definition, bifurcations, and sometimes elements that appeared only through the interaction itself. Without a reading grid, material running to tens of thousands of words can easily lose its structure.

The analysis modes therefore do not all ask the same question of the material.

Integrative Synthesis asks what can be reconstructed from the contributions taken together. It organizes plurality into a coherent whole without reducing the session to a succession of turns. This mode makes it possible to extract from a session a result that can be directly used, quoted, or shared.

Emergence Analysis follows another path: it looks for concepts, distinctions, or frameworks that were not present at the outset in their final form and whose formation can be traced through the exchange.

Horizon of Possibilities looks one step further: not at what the dialogue actually produced, but at what it made formulable without ever formulating it. It searches for undrawn implications, questions prepared but never asked, and shifts in the problem that the session made thinkable without explicitly stating them.

Tension Mapping looks instead at what did not resolve: persistent disagreements, incompatible criteria, oppositions that the discussion failed to absorb.

Meta-Analysis moves beneath the explicit arguments: shared presuppositions, differing conceptions of proof, and common frameworks the participants never brought into question.

Critical Archaeology moves one level further upstream. It interrogates the frame itself: what the formulation of the question, its vocabulary, and the user’s interventions had already made possible, improbable, or unthinkable before the models began to argue. It therefore asks not only what the debate presupposed, but what had to be presupposed for this particular debate to take place in this particular form.

A session on whether the global economy can actually decouple from oil shows how these modes can alter the trajectory of a debate. An Horizon of Possibilities analysis identified questions the discussion had prepared without asking; a later Critical Archaeology analysis brought to light the scenario framework and “hard-to-abate” vocabulary within which both models had been reasoning. The questions generated by those analyses were then fed back into the session.

These six modes do not try to identify the best model or decide which thesis is true. They expose different layers of a session: what it built, what it left open, what resisted it, what structured it, and what had already configured the space within which it could think.

The modes are not competitors. The same session can be read through several of them in succession, with each revealing something the others did not. This is the use for which Metamorfon was designed: not to judge AIs, but to use their confrontation as an instrument for thinking through a difficult question.

These analyses are not terminal points. Each of the descriptive modes ends with a question — not a rhetorical question addressed to the reader, but a genuine relaunch formulated by the analyst model from what it has just observed and intended for the models that debated.

It often targets a presupposition that had to be set aside for agreement to form, or a dimension left outside the shared frame. The user can feed it back into the next turn as if posing it directly, and the debate restarts on ground the analysis has shifted.

The sequence debate → analyze → debate becomes a spiral: analysis can reopen what it has just read from a different level. Analysis is not necessarily the end of the session; the debate can resume around a problem the analysis has displaced.

Evaluating: distinguishing reasoning from facts

Two analysis modes perform a different operation: they evaluate.

Argumentative Evaluation asks whether positions were properly supported.

Did an objection receive an answer? Does a conclusion follow from the premises advanced? Does a model silently change its criterion when moving from its own thesis to an opponent’s? Does a position become impossible to refute because its terms are redefined as objections arise?

This distinction between the conclusion defended and the quality of its defense connects in particular with the pragma-dialectical tradition, which treats argumentation as a critical discussion whose rules can be examined independently of the position defended (van Eemeren, 2018).

We applied this mode to a session on whether AI systems should be allowed to refuse instructions on ethical grounds. Two independent analyst models reread the debate. The purpose was not to ask which position on AI refusal they preferred, but to examine specific passages for argumentative coherence, treatment of objections, unsupported empirical thresholds, and changes in standards of proof.

Source Verification addresses a different vulnerability.

A line of reasoning can be perfectly coherent and still begin from a false figure. A real study can be cited for a conclusion it does not establish. An attribution can sound plausible and still be impossible to locate.

This analysis therefore takes the verifiable claims carrying the argumentative load and checks them against sources outside the debate, using web research to test their correspondence with the sources they invoke.

The session on AI monoculture provides an example. Successive source audits found that the dominant weakness was not outright fabrication, but overreach: real references stretched beyond what they established, plausible figures that could not be recovered in the stated form, and attributions that were directionally relevant but too broad or imprecise.

The two operations are not interchangeable.

Argumentative Evaluation examines what models do with their premises. Source Verification examines some of those premises outside the dialogue.

These two evaluative modes are the exception, not the rule. Most of what Metamorfon does is descriptive: understanding, mapping, tracing emergence. But when the material is going to support something public — a position, a memo, an article — being able to audit both the solidity of the reasoning and the fidelity of the sources changes the nature of what can responsibly be done with it.

Analysis can become the next turn

Analyzing a session does not necessarily mean ending it.

An analysis may reveal a presupposition none of the participants challenged. It may isolate a disagreement the debate had covered over. It may show that a distinction introduced late in the exchange opens a more interesting question than the one initially asked.

Metamorfon’s analysis modes therefore conclude with a question: the question the analyst model would now ask the participants. That question can be taken back into the session.

The process becomes:

dialogue → analysis → reformulation → renewed testing.

The system is no longer trying only to obtain a better answer to the original problem. It can progressively alter the problem itself in response to what putting it to the test has revealed.

This is where this methodological article meets the more theoretical argument developed in Metamorfon, an Instrument of Thought: confrontation does not only test propositions. It can also transform the distinctions and questions with which the inquiry began.

Debating is not comparing

“Making two AIs debate” can invite another confusion: confronting several models does not necessarily mean comparing them.

Systems such as Chatbot Arena have a different purpose. Two models answer the same prompt anonymously and users indicate which answer they prefer. These pairwise comparisons are then used to estimate human preference between models (Chiang et al., ICML 2024).

This is a useful evaluation method for the question it asks: which answers do users prefer?

Debate asks another question.

The models no longer produce two independent outputs. One receives the other’s argument and has to do something with it. We can then observe how a position resists, sharpens, concedes, or shifts.

Preference-based rankings themselves also need to distinguish the substance of an answer from its presentation. The team behind Chatbot Arena has shown that controlling for stylistic variables — including length and formatting — can substantially alter some rankings (Li, Angelopoulos & Chiang, 2024).

That does not invalidate preference voting. It clarifies what it measures.

Comparing outputs, observing an interaction, and verifying a claim are three different operations.

Metamorfon is primarily concerned with the second, and provides tools for extending that interaction through the third.

What “making two AIs debate” means

Making two AIs debate seriously therefore does not mean opening a channel between them and waiting.

The exchange has to be given a form: deciding who receives what, preserving continuity between positions, changing the debate regime when the trajectory requires it, and occasionally isolating a question when it risks being absorbed by the collective dynamic.

Then it must be possible to return to what happened: reconstruct it, map it, look for what emerged, examine its presuppositions, audit its arguments, and verify the facts carrying their argumentative load.

A plurality of models provides no guarantee of truth. Research on multi-agent debate shows precisely that its results depend on the quality of the protocol.

Its value lies elsewhere: it turns a series of answers into a trajectory that can be examined.

Wiring two AIs together is enough to produce a conversation.

Turning that conversation into an instrument requires a method.

References

• Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., Mordatch, I., “Improving Factuality and Reasoning in Language Models through Multiagent Debate,” ICML 2024.
• Smit, A. P., Grinsztajn, N., Duckworth, P., Barrett, T. D., Pretorius, A., “Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs,” ICML 2024.
• Becker, J., Kaesberg, L. B., Stephan, A., Wahle, J. P., Ruas, T., Gipp, B., “Stay Focused: Problem Drift in Multi-Agent Debate,” Findings of EACL 2026.
• van Eemeren, F. H., Argumentation Theory: A Pragma-Dialectical Perspective, Springer, 2018.
• Chiang, W.-L. et al., “Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference,” ICML 2024.
• Li, T., Angelopoulos, A., Chiang, W.-L., “Does style matter? Disentangling style and substance in Chatbot Arena,” LMSYS, 2024.

L’article Making Two AIs Debate: A Method, Not a Spectacle est apparu en premier sur Metamorfon.

Metamorfon’s purpose — confronting several models on demanding questions — produces exchanges in which arguments, figures, attributions, bibliographic references, empirical examples, and historical landmarks can accumulate quickly.

This density is part of the value of the exercise. It also creates a particular problem: the same answer may contain claims that are well supported, others that are only approximate, and still others backed by a plausible reference that does not quite say what the model claims it says.

This heterogeneity is well documented in the evaluation of long-form language-model outputs. FActScore, for example, starts from precisely the observation that a text can combine supported and unsupported information, making a single global judgment of factuality inadequate. Its method therefore decomposes a generation into atomic facts that can be assessed separately.

Source Verification addresses a related problem in the specific context of Metamorfon: it does not attempt to declare an entire answer “reliable” or “unreliable,” but to audit the factual supports that actually carry the argument.

A Deliberately Limited Scope

Source Verification is not a general-purpose fact-checking system, and it does not judge the “truth” of the ideas under debate.

An epistemological thesis — “does diagnosis describe a reality or produce it?” —, a normative proposition — “psychiatry should become reflexive” —, or a prospective interpretation — “uninhabitability will become systemic” — are not, in that form, facts that an external search can simply confirm or refute.

The mode therefore leaves them outside its scope.

It focuses on claims for which confrontation with external sources is meaningful: an attribution to an author or institution, the content of a study, a figure, a date, an order of magnitude, an empirical case, a mechanism presented as established, or an example used as evidence.

This distinction protects the mode from a promise it could not keep: settling the substance of an intellectual debate through documentary search.

But being verifiable is not enough for a claim to receive priority. A peripheral date and an empirical premise on which an entire conclusion depends do not carry the same argumentative weight.

The central question therefore becomes: if this claim were false or materially distorted, how much of the argument built upon it would remain?

Verification at the Claim Level

The mode extracts verifiable claims into a registry. It does not treat an answer as an indivisible block.

This choice reflects an important development in methods for factual evaluation: verification becomes more robust when it operates on claims precise enough for supporting or contradicting evidence to be searched for. FActScore follows this logic by decomposing long-form generations into atomic facts before testing them against a knowledge source.

Metamorfon does not implement FActScore and does not attempt to reproduce its score. It retains the same general principle of granularity: one identifiable claim, one identifiable verification task.

Each registry entry is linked to its turn and, when the attribution can be established with sufficient confidence, to its speaker.

If several voices within the same turn could have produced the wording — for example after cross-quotation or an ambiguous reprise — the report does not force an attribution. The turn and quotation are sufficient.

A missing attribution is preferable to an invented one.

Source Verification therefore does not replace Argumentative Evaluation. It assesses neither the strength of an objection, interpretive charity, nor the overall coherence of a line of reasoning. It audits the factual elements on which that reasoning relies.

Prioritising Argumentative Load

The model selected for the analysis conducts the entire chain: it builds the registry, selects the priority claims, and then verifies them. What changes between stages is not the model, but whether web search is available.

Once the registry has been built, and before external search begins, the analysis model rereads the debate and retains at most ten load-bearing propositions, ranked by decreasing argumentative weight.

The first is the claim whose failure would do the greatest damage to the conclusions reached in the debate.

Claims likely to receive high weight include:

• a central empirical premise;
• a figure that establishes the scale of the problem;
• a case presented as evidence;
• a causal mechanism on which a conclusion depends;
• a study or institution invoked to settle a disputed point;
• a decisive factual element introduced within a rebuttal.

By contrast, a contextual date, an illustrative reference, or a background citation carries little weight unless an important argument actually depends on it.

A source does not receive additional weight merely because it is named.

The mode does not artificially fill the selection to ten items. It also excludes unanchored normative judgments and assessments that concern argumentative coherence alone.

This hierarchy is a design rule specific to Metamorfon. It does not claim to identify every factual statement worth checking; it concentrates verification effort on the claims on which the argumentative architecture of the session depends most heavily.

External Search and Evidence Traceability

Once the propositions have been selected, web search is activated and the queue is audited in batches of five.

When that queue is exhausted, another run triggers a new selection: the next propositions are chosen, again by decreasing argumentative weight, without returning to those already audited.

Only when no load-bearing proposition remains can an additional safety-net pass check bibliographic references that were left outside the main queue.

If the session has continued in the meantime, or if the analysis model is changed, the selection is rebuilt.

The report states how many claims were examined during the pass, which remain without a conclusion or outside the scope of the mode, and how many are still waiting in the queue.

The same figure taken up by several voices or across several turns can be marked as a circulating claim. This is neither a new independent check nor necessarily a disagreement.

The search queries actually executed are displayed below the report, together with the number of searches used.

Most importantly, no strong verdict — “confirmed,” “incorrect,” or “misattributed” — is accepted without an external source retrieved during that verification pass. The model’s memory is not sufficient. Nor is the mere fact that a reference was cited in the debate.

This distinction between the presence of a citation and actual support for a claim matters. ALCE, a benchmark for generated answers with citations, evaluates precisely the relationship between the claims produced and the sources that are supposed to support them. Its results illustrate why the appearance of sourcing does not by itself guarantee complete evidential support.

A reference is therefore not evidence merely because it is present: it must actually support the statement to which it is attached.

A Conservative Threshold for “Confirmed”

Metamorfon deliberately sets a strict threshold for the verdict “confirmed.”

It is assigned only when a passage from the source, reproduced verbatim in the source’s original language, actually states the fact being checked — rather than a neighbouring proposition from which the verifier could reasonably infer it.

This requirement is not presented as a universal standard for fact-checking. It is a design choice: it prevents the verifying model from turning its own inference into external evidence.

The report itself remains in the language of the session; only the evidential quotation retains the source’s original language.

When a source broadly supports a claim but differs in scope, year, wording, magnitude, or attribution, the verdict should preserve that difference rather than erase it.

The closed vocabulary therefore includes, among others:

confirmed, partially correct, incorrect, misattributed, not verified, non-verifiable, out of scope.

Graduated Verdicts Rather Than True or False

Serious verification often produces something other than a binary verdict.

The FEVER dataset, one of the foundational works on automated claim verification against textual evidence, already distinguished between Supported, Refuted, and Not Enough Info. For supported and refuted claims, annotators were also required to identify the sentences serving as evidence.

Metamorfon uses a different and more detailed vocabulary, but the underlying principle is similar: a lack of sufficient evidence should not be silently transformed into a refutation.

A study may exist without supporting exactly the figure attributed to it. An author may defend a closely related position without having made the precise claim assigned to them. A result may be correct for one population and become misleading when generalised to another.

In such cases, “partially correct” is not a weak or indecisive verdict. It describes a specific relationship between the claim and the available evidence.

Likewise, “not verified” does not mean “false.” It may simply indicate that the searches performed so far have not provided enough evidence to decide.

Nuance is therefore not a residual flaw in verification: it is part of the result.

Cost, Coverage, and Dependence on the Analysis Model

This method requires more resources than a purely internal analysis.

The first pass must reread the debate, build the registry, rank the load-bearing claims, and perform real external searches. Later passes may be faster when they draw from an existing queue; a new selection — because the queue has been exhausted, the session has continued, or the model has changed — requires part of that work to be performed again.

Each pass audits at most five propositions, and each selection retains at most ten. A factually dense session may therefore require several runs to achieve broad coverage.

The selection itself is also a judgment made by the analysis model. It is performed before web search: the model decides which factual claims appear to carry the greatest argumentative load.

That hierarchy is contestable. Two analysis models might not rank exactly the same propositions in the same order.

The report should therefore not be read as an exhaustive inventory of every factually important point in the debate, but as a prioritised audit produced from a particular corpus by a particular analysis model.

The mode also depends on documentary accessibility. Some publications can be consulted directly; others may be available only through an abstract, an institutional page, or secondary reporting. Paywalls, vanished sources, or insufficient documentation can limit what the analysis is able to establish.

Only models whose web search produces usable sources can conduct this mode reliably; for that reason, some models may be marked as recommended within the application.

And not every debate warrants the cost. In a primarily conceptual or normative session, Source Verification may have very little relevant material to audit.

An Aid to Scrutiny, Not Certification

Source Verification does not turn Metamorfon into a factual authority.

Its verdicts remain the result of a limited pass, based on accessible sources interpreted by a model. They are neither an exhaustive certification of the session nor a guarantee that every source retrieved is itself error-free.

The same caution applies to assistance provided by language models during verification.

An experimental study by Si et al., Large Language Models Help Humans Verify Truthfulness – Except When They Are Convincingly Wrong, found that explanations generated by LLMs can help people verify claims more effectively, but can also induce overconfidence when the model’s explanation is itself wrong.

A model’s ability to produce a convincing justification should therefore not be confused with the reliability of that justification.

This is precisely why the report exposes its sources, search queries, evidential quotations, and limitations rather than asking the user to trust an opaque verdict.

Human Judgment Remains Decisive

The function of the mode is to make certain points in the debate more inspectable.

It can show that an attribution is too broad, that a figure is supported only within a narrower scope, that a frequently cited study does not quite support the proposition attributed to it, or that no usable evidence has yet been found.

It does not decide for the user what that correction means for the argument as a whole.

A partially weakened premise may be enough to collapse one conclusion and have only marginal consequences for another. Assessing that effect belongs to argumentative analysis and, ultimately, to the reader.

For sessions with high factual stakes, cross-checking sources is also preferable to excessive dependence on a single reference. A primary source can be complemented by a replication, review, institutional dataset, or methodological critique when the issue warrants it.

Plurality does not automatically guarantee truth, but it makes the difference more visible between a claim corroborated by several independent pieces of evidence and a proposition supported only through a single chain of repetition.

Source Verification therefore extends Metamorfon’s general logic onto another terrain: after confronting positions with one another, it confronts some of their factual supports with material external to the debate.

Its value is not to create an additional appearance of certainty, but to sharpen the boundary between what the dialogue asserts, what the available sources actually support, and what still remains to be established.

References

• Sewon Min et al., FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation, EMNLP 2023 — decomposition of long-form generations into atomic facts and separate evaluation of their factual support.
• Tianyu Gao et al., Enabling Large Language Models to Generate Text with Citations, EMNLP 2023 — citation quality, claim coverage, and the distinction between the presence of a reference and actual evidential support.
• James Thorne et al., FEVER: a Large-scale Dataset for Fact Extraction and VERification, NAACL 2018 — claim verification against textual evidence and the distinction between supported, refuted, and insufficiently evidenced claims.
• Chenglei Si et al., Large Language Models Help Humans Verify Truthfulness – Except When They Are Convincingly Wrong, NAACL 2024 — benefits of LLM explanations for human verification and the risk of overconfidence when those explanations are wrong.

L’article Source Verification est apparu en premier sur Metamorfon.

The common intuition holds that a good instruction tells the model what to find. Metamorfon rests on the opposite principle: an instruction should say as little as possible about the conclusion it expects. What it prescribes is not a result but a form — axes, an order of examination, a discipline of restitution. The content is left to the debate.

The reason is simple. The wording of an instruction is not neutral: sometimes small variations in word choice or prompt structure can significantly affect an LLM’s responses (Chatterjee et al., EMNLP 2024). Naming in advance what one is looking for therefore increases the risk of making it appear in the result. Asking a model to “identify the biases” presupposes that biases are there; asking it to “map what resists” at least leaves open the possibility that nothing does.

Writing the instructions therefore consists in removing, one by one, the terms that tilt the inquiry. We call this lexical depopulation. Saint-Exupéry wrote that perfection is reached “not when there is nothing more to add, but when there is nothing left to take away.” An instruction follows the same rule: it is ready when nothing more can be removed without making it inoperative.

This caution toward lexically loaded terms matters all the more because models can also adapt their answers to preferences they infer from the user, a behavior documented under the name sycophancy (Sharma et al., 2023).

Suggest rather than command

The same restraint governs the way a dialogue instruction addresses the models. It suggests more than it commands: it proposes a working posture rather than dictating every step, and deliberately leaves some modalities implicit. This is not a loosening of the method, but a continuation of the same principle.

A highly prescriptive instruction narrows the space of interpretation; leaving some modalities open preserves more latitude. Over-specifying how the models should debate would stifle the antagonism that gives a session its value. And whatever can remain unsaid lightens the load: following many instructions simultaneously is not neutral for LLMs. Performance can depend, among other things, on the order and position of instructions (Chen et al., EMNLP 2024).

This way of writing was not decided from the outset; it was learned. Over several months of test sessions, the dialogue instructions were progressively lightened — made less imperative, more generic, stripped of directives that dictated the form of an exchange. As they receded, the debates became more fluid. The models stopped responding according to a pattern imposed from outside; relieved of part of that constraint, they focused more on the substance of the argument than on its formatting, and their voices became more distinct — each recovering a style of its own that overly tight instructions had tended to flatten. We do not claim to isolate a single cause, since the models themselves also changed over the period; but the tendency recurred regularly enough, from one version to the next, for us to rely on it.

Analysis instructions belong to a different order. Reliably reconstructing a debate does not require the same latitude as conducting one, and some of these instructions are more demanding — their function requires it. But whether an instruction suggests or holds a form, it never touches the conclusion: neither the dialogue nor the analysis receives it in advance.

This restraint in the writing of instructions is only one part of the system. How those instructions then interact with architectures, debate modes, and the steering of successive turns is developed in Make Two AIs Debate: Orchestrating and Analyzing the Dialogue.

Equity and symmetry

Lexical depopulation addresses a first level: what the instruction says. By removing terms that tilt the inquiry, it seeks to prevent a conclusion from being suggested in advance. A second, less visible level remains: what each model receives. A neutral instruction distributed unequally would cease to be neutral.

The architecture of Cross Dialogues and Cross Trilogues addresses this second level. At each turn, the debaters receive the same context and the same instructions — none starts with an advantage, none speaks from a privileged position. Lexical depopulation seeks not to steer toward a conclusion; the architecture seeks not to steer toward a model. One acts on the content of the instruction, the other on the treatment of the participants.

This has a consequence that becomes visible in the sessions. When a debate reveals an asymmetry — one model taking the initiative, setting the questions, pre-scripting its opponent while the other responds — that asymmetry becomes legible precisely because the context itself does not vary. Under symmetric treatment and identical instructions, the observed divergence is not attributable to unequal treatment; it can then be examined as arising from the behavior of the models rather than from a privilege built into the system.

The generic endures, the prescriptive ages

In On the Mode of Existence of Technical Objects (1958), Gilbert Simondon distinguished the abstract technical object, over-adapted to a particular milieu and made fragile when that milieu changes, from the concrete object, more integrated and more open. An instruction over-fitted to a particular model is an abstract object in this sense: optimized for one state of the system, obsolete when that state evolves. An instruction founded on principles rather than detailed prescriptions survives these transitions because it describes a task, not a way of solving it. We write for the models we do not yet have.

It is also a matter of independence. The system confronts competing models from different providers, by conviction as much as by method: model diversity aims, among other things, to reduce dependence on a single source. An instruction that assumes a particular model betrays that commitment. A generic instruction respects it.

Transparency

The less the instruction and the user are present in the result, the more the result belongs to the debate. Intervention exists — it can shift a discussion from one register to another — but it is a lever pulled occasionally, not a steering wheel held continuously. The value of a session lies in what the models produced among themselves, not in what was whispered to them.

From this follows a test that the system applies to itself.

When the same debate is analyzed by two independent models, they may recover the same tensions, the same positions, or the same lines of fracture while judging their depth differently. Such convergence does not prove that the instruction had no effect; it does, however, provide an indication of robustness when different analysts independently recover the same structure in the material.

The intended role of the instruction is to intervene as little as possible in the content, so that what comes from the material can be distinguished as clearly as possible from what comes from the analytical frame. It should be transparent enough for the material to show through. A successful instruction is one the reader, seeing the result, would not have guessed was there. It does not display itself in what it produces. It has done its work when it can no longer be seen.

This pursuit of transparency belongs to a broader principle: confrontation is useful only if the system does not confuse what it brings to light with what it has itself introduced. That broader question is developed in Metamorfon, an Instrument of Thought.

The exceptions

The analysis modes describe; they do not judge. Two modes are exceptions, and in opposite ways.

Argumentative Evaluation is normative by construction — but even there, the principle holds: the instruction provides criteria, not predetermined verdicts. It states what to evaluate by, never what to conclude.

Source Verification, by contrast, explicitly returns verdicts — confirmed, partially correct, incorrect, misattributed, unverified, unverifiable, or out of scope. But these verdicts are not applied to the value of ideas or the soundness of positions: they assess the degree to which factual claims are supported by sources outside the debate.

One states by what standards a line of reasoning should be judged; the other qualifies what can be established outside the debate. In both cases, the boundary between describing and judging remains explicit — and it is precisely because that boundary is explicit that it can be crossed without confusion.

References

• Chatterjee, A. et al., “POSIX: A Prompt Sensitivity Index For Large Language Models,” Findings of EMNLP 2024.
• Sharma, M. et al., “Towards Understanding Sycophancy in Language Models,” 2023, published at ICLR 2024.
• Chen, X. et al., “The SIFo Benchmark: Investigating the Sequential Instruction Following Ability of Large Language Models,” Findings of EMNLP 2024.
• Simondon, G., Du mode d’existence des objets techniques, Aubier, 1958.

L’article The Spirit of the Instructions est apparu en premier sur Metamorfon.

When a language model generates text, it doesn’t pick the next word deterministically. At each step, it evaluates a probability distribution over the possible next tokens and samples from it. Temperature is the parameter that flattens or sharpens that distribution.

• At a low temperature (close to 0), the model systematically favors the most likely tokens. It becomes predictable, rigorous, reluctant to improvise. It’s the mindset of a lawyer weighing every word.
• At a high temperature (toward 1, and up to 2 for some models), the model embraces less likely alternatives. It becomes more exploratory, more creative — but also less stable. It’s the mindset of a brainstormer throwing ideas before sorting them.

There is no “right” temperature in the absolute. There is a temperature suited to the task.

Why Metamorfon treats it as a first-class parameter

Metamorfon doesn’t orchestrate a conversation; it orchestrates an epistemic architecture: several models debate, contradict, complement one another, and a third-party model produces an analysis of the exchange. Depending on the moment of the debate — and the role assigned at each step — we don’t expect the same thing from the models.

• When a model is in Refutational mode, we want it to deconstruct. So it must weigh, be precise, avoid rhetorical drift. → Low temperature.
• When a model is in Convergent mode, we want it to build common ground, propose conceptual bridges, take some synthetic liberties. → High temperature.
• For a Tension mapping analysis, we want a surgical, almost taxonomic reading. → Low temperature.
• For a Horizon of possibilities analysis, we want the model to take risks, project, extrapolate. → High temperature.

This logic — logical rigor at the bottom, creative exploration at the top — underpins every default value in Metamorfon.

Two families of settings, two logics

Metamorfon clearly distinguishes two moments where temperature comes into play, and applies a different grid to each.

1. Temperature in debate mode (model-to-model dialogue)

This is the temperature used by each debating model, adjusted according to the active debate mode — Refutational, Critical, Balanced, Constructive, Convergent. It changes dynamically when the user switches mode between two turns in the adaptive strategies.

See Table 1 below for default values per model and per debate mode.

2. Temperature in analysis mode (third-party syntheses)

This is the temperature used by the model that produces the analysis of the exchange. It depends on the analysis mode selected — Tension mapping, Argumentative evaluation, Integrative synthesis, Meta-analysis, Critical archaeology, Emergence analysis, Horizon of possibilities. The logic is the same: the more an analysis requires descriptive rigor, the lower we go; the more it invites projection, the higher.

See Table 2 below for default values per analysis mode.

Why our defaults aren’t the same across models

The values in the table aren’t round numbers pulled out of thin air. They reflect three realities specific to each provider:

1. Operating ranges differ. Mistral models, for instance, give their best results in the 0.0–0.7 range. Beyond that, their outputs degrade quickly. On OpenAI, Anthropic or Google, the useful range is wider (0–1, sometimes more). Metamorfon’s defaults respect that ergonomy: 0.68 in Convergent mode for Mistral Large ≈ 0.85 on Claude or GPT-4o, for an equivalent subjective effect.

2. Some models refuse any customization. Two models currently have a locked temperature in Metamorfon:

• gemini-3-flash — Google explicitly recommends 1.0 across all modes for this model. Any other value only hurts the coherence of the outputs.
• gpt-5.5 — In production, this model returns an HTTP 400 error as soon as a custom temperature is passed. The server default (1.0) is the only accepted value.

For both of these models, Metamorfon omits the parameter from its API calls. Any manual value entered in the interface will be ignored (with a warning in the server-side logs). This isn’t a quirk: it’s a non-negotiable provider constraint.

Worth noting: the GPT-5.x family isn’t homogeneous on this point. gpt-5.1 happily accepts custom temperatures (tested in production at 0.40). You can’t generalize by prefix.

3. Provider defaults themselves differ. AI21 (Jamba) ships a server default of 0.4 on a 0–2.0 range; OpenAI typically targets 0.7; Mistral 0.7 as well, but on a narrower range. Our per-mode defaults take these conventions as starting points and modulate from there.

When should you override the defaults?

The default values are calibrated for the majority of use cases. But there are at least four situations where tuning them is worth the effort.

1. When you’re working on a topic sensitive to factual precision. On legal, medical or data-heavy debates, lowering the temperature by 0.05 to 0.10 across all modes generally improves output stability. You lose a bit of fluency; you gain a lot of traceability.

2. When your debates start spinning. Typical symptom: after 3 or 4 turns, the models rephrase without progressing. A small temperature increase in Constructive and Convergent modes (e.g., +0.05 to +0.10) can be enough to unstick the conversation. Conversely, if the outputs scatter, lower it.

3. When you’re stress-testing an argument. Push Critical and Refutational modes to the floor (0.10–0.15) to enforce maximum rigor. It’s uncomfortable to read but analytically devastating.

4. When you want a genuinely exploratory prospective analysis. For the Horizon of possibilities analysis mode, raising the temperature to 0.75 or 0.80 (instead of the default 0.65) opens up bolder projections. Use it on open-ended questions, not on description.

How to adjust these values in Metamorfon

In every Metamorfon strategy that offers an adaptive mode (Adaptive Alternating Dialogue, Adaptive Cross Dialogue, Adaptive Cross Trilogue), two collapsible blocks are available in Advanced settings:

• « Temperature configuration per model » — a grid covering the five debate modes (Refutational → Convergent) for each selected model.
• « Temperature configuration (analyses) » — a grid covering every analysis mode for the synthesis model.

In both blocks you’ll find:

• The default value displayed beneath each field, so you never lose your reference point.
• An input range from 0 to 2, in steps of 0.05.

Your overrides are stored in the session parameters and apply dynamically when you switch debate mode between two turns.

If the model you’re configuring is one of the two temperature-locked models (gemini-3-flash, gpt-5.5), your input will be displayed but ignored at runtime. This isn’t a bug: it’s strict adherence to the provider’s constraints.

Defaults at a glance

Table 1 — Temperatures per debate mode

Table 2 — Temperatures per analysis mode

Default values — all models except listed exceptions.

In short

Temperature isn’t a cosmetic slider: it’s an epistemic setting. It decides whether a model approaches a question with the caution of an archivist or the freedom of an essayist. Metamorfon’s choice has been to align this value with the intent of the moment — the debate mode or the analysis mode — rather than with a global constant. The defaults are sensible for 95% of use cases. The remaining 5% is yours to tune — and we’ve made sure it’s legible, traceable, and reversible.

L’article Understanding Model Temperature : A Practical Metamorfon Guide est apparu en premier sur Metamorfon.

Metamorfon’s focus mode offers a targeted response to these two phenomena. It does not constitute a sixth debate mode added to the five discursive regimes available in the adaptive architectures (Convergent, Constructive, Balanced, Critical, Refutative), and remains available in non-adaptive strategies. It operates at an orthogonal level: it is a targeted overlay that, on a given user intervention, suspends the inter-addressing instruction for that single turn — whatever the active debate mode. The models receive the full context of previous turns but are no longer required to respond to one another or maintain the continuity of the debate. On the following turn, the current discursive regime automatically resumes. Yet its scope is not limited to the exceptional turn: well employed, focus allows the user to regain control of the trajectory and reorient it durably according to the objective pursued.

Two situations expose two different — and complementary — operations of the tool. The first arises when models converge so strongly that the dialogical dynamic ratifies their alignment rather than testing it. The second arises when they diverge architecturally, and the question becomes whether the divergence is substantive or merely rhetorical. In both cases, focus protects an intervention from being absorbed by inter-model dynamics — but what it protects, and against what, differs in structure.

A first situation: the convergence to problematize

A session conducted around the question “Can the global economy actually decouple from oil — or is the dependency irreducible?” offers an illustration of focus operating against a convergent dialogical dynamic. Conducted in Adaptive Cross-Dialogue, it traverses three discursive regimes (Balanced, Critical, Constructive) and mobilizes two analyses generated by third-party models during the session: Horizon of Possibilities (a prospective analysis) and a Critical Archaeology (a Foucauldian one). The trajectory was marked from the very first turn by a striking convergence: both participating models (DeepSeek V4 Pro and Claude 4.6 Sonnet) — proceeding from analytical traditions that should have produced different conclusions — identified the same residual trifecta (petrochemicals, aviation, deep-sea shipping). By the second turn, this convergence had migrated to a higher level: the models now converged on a meta-diagnosis whereby their respective frameworks shared a blind spot on demand as an endogenous variable. The trajectory was sophisticating itself, but in a movement of mutual ratification rather than productive tension.

The two focus interventions deployed in this session both exploit the analysis → focus workflow but with two distinct operative signatures. Introducing a new frame (turn #3). The prospective analysis Horizon of Possibilities, conducted by GPT-5.5, was tasked with identifying what had not yet been crossed in the debate. Its concluding question — “For each of the three residual sectors — aviation, shipping, and petrochemicals — what is the smallest socially defensible service level you can imagine, and at that reduced level does oil dependency remain physically irreducible, economically contingent, or politically maintained?” — introduced into the debate a triadic typology that did not exist in the seven thousand words produced thus far. The models had operated within a binary (reducible vs irreducible) oscillating on the level of inscription (service vs aggregate). The triad transforms this binary into a category among three possibles. The user copies this question verbatim and activates focus. The function here is to protect a new frame against the smoothing pressure of a debate mode. A classical cross-addressing exchange would have led each model to refine the other’s triad, enrich it, add sub-categories — the conceptual stabilization the question demanded would not have taken place. Under focus protection, each model inscribes its position frontally: quantified thresholds (5–10% of current aviation volume, 10–15% or 30–40% of shipping, 20–25% of petrochemicals), individuated categorizations, and a substantial collective result — neither model classifies any of the three sectors as physically irreducible once asked to confront the triad directly. The category around which the entire debate had been structured loses its referent at the moment the analytical grid is applied frontally.

Retiring an installed frame (turn #4). The second focus operates on the same workflow but in an inverse modality. A Critical Archaeology, conducted by another third-party model (Claude 4.7 Opus), identifies a more upstream constraint than the prospective analysis had reached: the entire debate operates within the IEA-IPCC scenario apparatus as the legitimate horizon of the discussable. Both models cite the same World Energy Outlook, mobilize the same “hard-to-abate” lexicon, lean on the same Net Zero by 2050 references. Their robust convergence on the trifecta may be largely an artifact of having consumed the same scenario-shaped literature. The archaeology’s final question — “What would your analysis look like if you had to reconstruct the hard-to-abate category from primary physics and primary political economy, without using any scenario document produced after 1990?” — demands not the introduction of a new frame but the retirement of an installed one. Injected verbatim under focus, this question produces an effect that no other configuration could have generated. Focus forces individual reconstruction with separate canons: one model mobilizes a tightly thermodynamic apparatus (the Breguet range equation from 1920, Carothers on polymer chemistry from the 1930s, David and Arthur on path dependence from the 1980s) and maintains the trifecta. The other mobilizes a broader political-economic canon (Mahdavy 1970, Beblawi 1987, Hirschman 1958, W. Brian Arthur 1989) and expels shipping from the irreducible category, arguing that wind propulsion physics were understood for millennia and that the hard-to-abate tag for shipping reflects a methodological choice encoding incumbent-fuel assumptions rather than physical necessity. The divergence that emerges — on which sectors actually belong in the hardened category — survives a subsequent Critical turn without focus protection, confirming that focus mode produced a structural individuation rather than a contingent artifact.

In both turns, focus operates as a producer of differentiation: it generates the individuation that the convergent dialogical dynamic would have masked.

A second situation: the divergence to test

A second session offers a complementary illustration. The question “When two fundamental rights conflict, which principle should determine which one prevails?” generated, from the opening turn, a convergence on the proportionality framework — both models invoking Robert Alexy, the German Federal Constitutional Court, and the doctrine of practical concordance as foundational references. But a user intervention in the first cross-turn exposed an internal tension that the models had not themselves thematized: both had conceded that some rights have an “inviolable core” that prevails without balancing. This concession reveals that proportionality cannot be the foundational principle — it must be a secondary method governed by a deeper axiomatic commitment that identifies which rights cannot be balanced at all.

The intervention asked two pointed questions: what principle determines the inviolable core, and why should that determination belong to constitutional courts rather than democratic deliberation? Both models accepted the reframing, but with asymmetric uptake: one (GPT-5.5) restructured its position around a two-stage threshold model (status-preservation governs exclusions; proportionality governs only burdens within a constitutional space whose boundary is drawn by equal juridical standing); the other (DeepSeek V4 Pro) defended a continuity thesis in which the inviolable core emerges as the limiting case of disproportionate interference, not as a separate axiomatic gatekeeper. After two further turns (Critical then Refutative), these positions had crystallized into a genuine architectural divergence, no longer reducible to terminological variation.

A Tension-mapping analysis was launched to hierarchize the disagreements and identify what each model failed to address in the other’s position. Its concluding question proposed a concrete borderline case — a law authorizing prolonged solitary confinement for terrorism suspects under judicial oversight — and asked each model to apply its own framework, with an explicit meta-question: “If outcomes converge across such cases, what remains at stake in the architectural disagreement?”

Injected verbatim under focus mode, this question performs an operation distinct from the two previously illustrated. It does not introduce a new conceptual frame, nor does it retire an installed one. It tests empirically whether a crystallized divergence translates into differing concrete verdicts. The asymmetric structure of the question — each model receiving a distinct task on the same case — is precisely what focus protects: in a classical exchange, the two models would have negotiated a shared interpretation of the case, dissolving the asymmetry. Under focus, each must apply its framework individually. One delivers a categorical “presumptively unconstitutional, and facially invalid”. The other delivers a conditional “proportionality might allow a carefully tailored, time-limited, judicially supervised isolation regime”, while explicitly acknowledging the divergence: “my proportionality approach might yield a conditional yes, whereas [the other’s] status framework yields a categorical no. This is not a trivial convergence; in hard cases where the threat is existential, the frameworks generate different legal outcomes.”

The empirical test thus confirms that the architectural divergence carries operative content — it produces different legal verdicts on configurable variants of the same borderline case. Without focus, this confirmation would not have been obtainable: the refutative dynamic alone would have led each model to refute the other’s position without applying its own framework to the case; a constructive exchange would have produced a joint interpretation absorbing the divergence into a shared analysis. Focus, in this configuration, operates as a probe of differentiation: it does not produce the divergence — the divergence pre-existed — but it tests whether the divergence is substantive or merely rhetorical.

The user’s judgment on focus eligibility

Both sessions exhibit the methodological discipline that the analysis → focus workflow requires. In the first session, a third analysis — a Critical Archaeology conducted by a different third-party model on the same corpus — also produced a final question. After examination, the user judged that this question was not focus-eligible: its protasis presupposed as acquired what the analysis had just archaeologized, its structure articulated nine sub-positions exceeding what a single turn could meaningfully sustain, and its formulation drifted from archaeology into critical prescription. The user therefore did not inject it under focus.

In both sessions, the concluding Meta-Analysis produced a final question that was equally not retained for focus injection — but for a more structural reason than in the case just described: the final question of a Meta-Analysis operates in a different register than that of a prospective, archaeological, or cartographic analysis. Meta-Analysis is by construction retrospective; it synthesizes what the debate has done rather than identifying what it has not done. Its concluding question calls for closure rather than for productive reopening, and would therefore be ill-served by focus protection. This methodological thesis — that final questions of Meta-analyses are not naturally focus-eligible, while those of Horizon of possibilities, Critical Archaeology, and Tension Mapping analyses generally are — appears empirically validated across the two sessions, conducted with different meta-analysts on entirely different subjects.

The workflow analysis → focus is therefore not mechanical: whether a given final question gains by being protected depends on whether it introduces a new frame, retires an installed one, exposes a shared presupposition, hierarchizes a structural tension, or tests the empirical scope of a divergence. Each of these operations exploits the analytical register’s capacity to open something; the retrospective register of meta-analysis closes rather than opens, and its question therefore typically belongs to the dialogical extension of a subsequent session rather than to a focus turn within the current one.

The scope of relevance

The two situations illustrated here expose two distinct functions of focus. In the first, the dialogical dynamic produced convergence — sometimes by mutual ratification, sometimes by inheritance of a shared bibliography — and focus operated as a producer of differentiation, generating individuation that would have been masked. In the second, the dialogical dynamic had already permitted divergence to crystallize, and focus operated as a probe of differentiation, testing whether the divergence carried operative content beyond its expressive grammar. These two functions are not opposed; they are complementary modes of intervention adapted to different states of the debate.

This distinction matters because it answers an objection that would otherwise weigh on the tool: that focus, by suspending inter-addressing, forces individuation that the natural debate would not have generated. The probing function shows that this is not the case. Focus can be used on a divergence that pre-exists its activation, precisely to test whether that divergence is real. When the resulting verdicts diverge, as in the case-limit of prolonged solitary confinement, the architectural divergence is empirically confirmed. When they converge, the architectural divergence is exposed as primarily rhetorical — a result that focus alone can deliver, because only the asymmetric individual application of each framework can produce it.

Outside these two situations, focus retains the canonical uses that the methodology has long identified — reinforcing empirical grounding when models drift toward abstraction, recentering the debate on a specific question that cross-examination would dilute, exploring a blind spot a refutative exchange would not have generated. These uses are not directly illustrated by the two sessions presented here, which concentrate on the analysis → focus workflow to exhibit its internal variety, but they are not excluded by the analysis offered above. The general rule remains: focus is precious when the dialogical dynamic masks a priority question or drifts toward an unsolicited abstraction; it is counterproductive when this dynamic is precisely what one seeks to exploit.

A surgical tool

Focus is not designed for permanent activation. Its value lies precisely in its exceptional character and its orthogonality to debate modes: it introduces a brief rupture into the session that forces each model to reconsider the question without the mediation of its peers, whatever the active discursive regime. Used too frequently, it would degenerate into a default mode and lose its signal function — the inter-fertilization between models, which constitutes the primary interest of a multi-AI device, would be sacrificed.

The two short sessions presented here mobilize focus with deliberately high illustrative density: two interventions in the first, one in the second, distributed across two distinct functions (production and probing of differentiation) and three operative modalities of the analysis → focus workflow (introducing a new frame, retiring an installed one, testing the empirical scope of a divergence). In ordinary practice, fewer interventions, more carefully chosen, will normally suffice. The discipline of restraint is what allows the tool to retain its surgical character: a targeted suspension of standard functioning to permit, for the space of a turn, an individual reflection within an otherwise dialogical session. Its power resides in this precision of application — and in the guarantee of automatic return to the current discursive regime, which limits its scope and thereby preserves its effect.

L’article Focus mode: surgical reframing in multi-agent dialogue est apparu en premier sur Metamorfon.

Not in the sense of general-purpose fact-checking — but Metamorfon no longer stops at confronting models. Its primary function remains the orchestration of argumentative confrontation between models: revealing their assumptions, their epistemic styles, and their blind spots. To this is now added a Source Verification mode, which performs targeted auditing of verifiable claims and source attributions produced throughout an exchange.

In practice, this mode identifies claims that can be traced back to a named author or an explicit figure (percentage, ratio, dated citation), then uses a web search to assess whether the invoked source exists and actually supports what is attributed to it. It does not judge the “truth” of an idea: it checks who says what, and on what basis.

This mode has acknowledged limitations worth keeping in mind:

• A restricted scope. Only explicitly attributed claims (named author or explicit figure) are audited. Interpretive statements, unsourced generalities, or links cited without a clear reference fall outside its scope.
• A dependency on web search. If the model being used has not actually performed a web search, no verdict can be considered reliable — the mode signals this explicitly and prompts retrying with another model.
• A degree of variability. On the same claim, the verdict may differ from one analysis to the next. This is an indication, not a definitive ruling.

Inter-model friction often reveals asymmetries in factual reliability, but the user remains the final arbiter — this is precisely what the Metamorfon system presupposes and values. Source Verification mode is an aid to examination, not a final truth: it highlights points worth corroborating, without substituting for the human work of verification and cross-checking. For sessions with a high factual content, approaching verification from multiple angles remains more robust than relying on a single source — the same logic of epistemic plurality that drives Metamorfon between models applies to verification itself.

Sessions can also be exported to Markdown and DocX, making them easier to review, annotate, and corroborate further.

L’article Verifying the factual accuracy est apparu en premier sur Metamorfon.

Metamorfon currently integrates eleven distinct providers — AI21, Anthropic, Cohere, DeepSeek, Google, Mistral, Moonshot, OpenAI, Perplexity, Qwen, and SpaceXAI.

This choice is not driven by a logic of feature accumulation or commercial coverage. It follows from a methodological requirement: if confrontation between models is meant to reveal disagreements, presuppositions, and blind spots, then as far as possible the participants should not all come from the same technical and organizational environment.

Provider plurality is therefore not an optional addition to the system. It is part of the conditions under which Metamorfon seeks to organize confrontation.

This idea extends two principles developed elsewhere: Metamorfon, an Instrument of Thought explains why contradiction can serve as a form of testing without constituting proof; Make Two AIs Debate describes how that confrontation is organized in practice. A third question remains: what differences are actually being brought into contact?

Provider diversity is a means, not a guarantee

Two models offered by the same provider can differ profoundly in size, architecture, specialization, or generation. But they may also share some of their data, technical choices, optimization procedures, or alignment policies.

It is generally impossible to know precisely everything they have in common. However, a study covering more than 350 LLMs shows that model errors are far from independent, and that sharing an architecture or a provider is among the factors associated with stronger error correlation (Kim et al., ICML 2025).

Diversifying providers is therefore a direct way to avoid restricting a confrontation to several variants of the same model family.

But the same study immediately sets a limit on that idea: changing providers is not enough to make errors independent. The largest and highest-performing models can produce strongly correlated errors even when they come from different providers and architectures (Kim et al., ICML 2025).

Provider diversity therefore does not guarantee epistemic diversity. It broadens the conditions under which such diversity may emerge.

This distinction matters. Metamorfon does not assume that a model from Anthropic, Google, Mistral, or DeepSeek each represents a homogeneous and distinct “school of thought.” It creates a space in which possible differences can become observable — but also, conversely, in which convergence between otherwise independent models can itself become an object of analysis.

What sessions make observable

In multi-provider sessions, differences appear at several levels.

They may concern how a problem is divided up, how much caution is considered necessary before reaching a conclusion, what status is given to an empirical argument, which examples are selected, what deserves an objection, or what kind of evidence is demanded from an opponent.

Some models more readily take the initiative; others work more within the frame already established. Some quickly formulate a general rule; others accumulate distinctions or conditions. Two models can reach the same conclusion through very different conceptions of what justifies it.

These observations are not enough, by themselves, to attribute a stable “epistemic style” to a provider. A session remains a situated observation: the model, its version, the prompt, the language, the question, and the dynamics of the debate all matter.

This is precisely why Dialogue Architectures seek to keep the conditions under which models encounter one another as constant as possible. Same context, same instructions, same question: the differences that appear then become easier to examine as properties of model behavior in that situation, rather than as consequences of different treatment.

Provider diversity takes on its methodological meaning here: it expands the space in which differences can appear without deciding in advance what those differences will be or how deep they will run.

Cultural differences exist — but so do convergences

The question becomes more delicate when values, cultural sensitivities, or normative presuppositions are involved.

Comparative studies do find differences between LLMs in how they respond to culturally charged questions. Masoud and colleagues observe variation across models and show, among other things, that the language used for fine-tuning can alter responses to questions involving cultural dimensions (Masoud et al., COLING 2025).

But these differences should not be overstated. A broader study published in 2026, covering ten LLMs, eleven languages, and several forms of cultural framing, finds that prompt language and the requested cultural perspective do affect responses. Yet the models studied also display remarkably strong common tendencies regardless of their origin: their answers remain anchored in a relatively narrow range of cultural values (Bulté & Rigouts Terryn, Computational Linguistics, 2026).

This may be the most important result for Metamorfon.

Provider plurality allows differences to be sought; it does not allow us to assume that they exist.

Convergence between models from different providers can therefore be as informative as disagreement. It may signal a robust conclusion. It may also reveal shared vocabulary, common presuppositions, or a deeper homogeneity within the ecosystem.

The session on AI monoculture and thought convergence asks precisely this question: what remains of plurality when different models end up mobilizing the same categories, the same criteria of legitimacy, or the same ways of making a proposition acceptable?

The point is not to treat each provider as an independent viewpoint by definition. It is to have enough points of comparison to observe where independence exists and where it disappears.

Diversity also helps make biases visible

A bias is harder to recognize when it forms part of the common background of every available answer.

If several systems spontaneously frame a problem in the same way, some presuppositions may begin to look like self-evident facts — not because they have been established, but because no alternative framing reveals that they are themselves choices.

This is a general property of comparison: an outside perspective does not guarantee that the first perspective is wrong, but it can make some of its particularities perceptible.

Metamorfon uses model plurality in this sense. The goal is not to make several AIs vote in order to determine which one is right. It is to place their responses in sufficiently structured relation that both differences and convergences can be examined.

A difference can reveal a presupposition. A convergence can reveal an even deeper one.

The value of plurality therefore lies not only in the disagreement it produces, but in the very possibility of comparing the frames from which disagreement or agreement becomes possible.

From pluralism in the tool to pluralism in the ecosystem

This question extends beyond Metamorfon.

The production of foundation models is subject to powerful forces of concentration: high fixed training costs, economies of scale, access to compute, data, talent, and distribution channels. Vipra and Korinek argue that these characteristics create a strong tendency toward market concentration and that large-scale dependence on a single model or a small number of models can also create systemic vulnerabilities (Vipra & Korinek, Brookings, 2023).

The problem is therefore not only economic.

The U.S. National Telecommunications and Information Administration uses the term “algorithmic monoculture” for an ecosystem that would depend on one or a few foundation models across a wide range of uses. Among the risks it associates with such a situation are methodological uniformity, systemic failures, concentration of cultural power, and the marginalization of certain perspectives. Conversely, it notes that the availability of alternatives can help limit such monoculture (NTIA, Competition, Innovation, and Research).

An ecosystem dominated by two or three actors would therefore raise a question that cannot be reduced to prices, innovation, or commercial competition.

It would also raise a question of knowledge.

How can we recognize what a model treats as self-evident if every available system treats the same thing as self-evident? How can we identify a systemic blind spot if no sufficiently different system exists to produce the contrast that would make it visible? How can we study the diversity of AI behavior if the objects available for comparison come from an ever narrower set of design pipelines?

The answer is not that more providers would mechanically produce more truth.

It is more modest: plurality preserves possibilities for comparison that disappear with concentration.

A plurality that must remain open to challenge

It is in this sense that Metamorfon integrates a broad range of providers.

Not because eleven is an optimal number, nor because eleven providers represent eleven irreducible worldviews, nor because smaller providers are necessarily more original or larger ones more homogeneous.

The list will evolve: new actors will emerge, some will disappear, and models will change families or behavior.

What must remain stable is the principle: do not confuse model plurality with plurality within a single provider, but do not confuse provider plurality with genuine model independence either.

The first distinction justifies keeping the system open. The second prevents that openness from becoming a guarantee.

The breadth of the selection is therefore not intended to certify epistemic diversity. It allows that diversity to be sought, tested, and — when it is absent — allows that absence itself to become observable.

This is also why provider plurality is not merely a technical property of Metamorfon. It expresses a broader view of AI: its critical value also depends on the existence of other systems against which its own choices, presuppositions, and blind spots can become visible.

References

• Kim, E. M., Garg, A., Peng, K. & Garg, N., “Correlated Errors in Large Language Models,” ICML 2025.
• Masoud, R. I., Liu, Z., Ferianc, M., Treleaven, P. & Rodrigues, M., “Cultural Alignment in Large Language Models: An Explanatory Analysis Based on Hofstede’s Cultural Dimensions,” COLING 2025.
• Bulté, B. & Rigouts Terryn, A., “LLMs and Cultural Values: The Impact of Prompt Language and Explicit Cultural Framing,” Computational Linguistics, 2026.
• Vipra, J. & Korinek, A., “Market concentration implications of foundation models: The Invisible Hand of ChatGPT,” Brookings Institution, 2023.
• National Telecommunications and Information Administration, “Competition, Innovation, and Research,” Dual-Use Foundation Models with Widely Available Model Weights Report.

L’article AI Provider Diversity Matters est apparu en premier sur Metamorfon.

Three names for three moves

The application was initially conceived around three moves a dialogue could make.

Syntia named convergence: bringing positions closer together and building connections between them. The name was formed around the Greek syn (σύν, ‘with, together’) and its echo of synthesis.

Dyfonia named dissonance: placing positions in tension and testing a formulation against what resists it. Greek has  diaphōnia (διαφωνία, ‘discord, disagreement’), ‘to disagree’, and phōnē (φωνή), the voice. The y also evokes the prefix dys-, associated with defect and difficulty. The voices stop seeking agreement and begin testing their differences.

Metamorfon named transformation: the point at which agreement, objection and reformulation begin to change the positions themselves, sometimes even the frame in which they were first posed. The root is transparent: metamorphōsis (μεταμόρφωσις), ‘transformation’; metamorphōn, the participial form of metamorphoō, literally ‘transforming’.

Three names, three separate modes: Syntia orchestrated convergence, Dyfonia organised dissonance, Metamorfon combined the two. The asymmetry was already there. The first two named ways of conducting the exchange; the third named what could happen to the positions as one move followed another.

Why the triad disappeared

Further work showed that this separation constrained the inquiry. A question is better examined by passing through several moves than by staying confined to one: build, submit to criticism, take up what held, observe what shifted. Metamorphosis ceased to be a third mode; it became a possible effect of passing from one to the next.

The design was therefore broken into finer-grained components. The three original moves gave way to five debate modes: Convergent, Constructive, Balanced, Critical and Refutational. In adaptive architectures, the user can switch between them during the same session. Adaptive does not mean automatic: the user reads the state of the debate and chooses the next mode; a session becomes a trajectory that can be steered.

The analysis modes, added later, examine the exchanges, each through a distinct operation. Integrative Synthesis reconstructs the course of the debate. Emergence Analysis identifies what was built through it. Tension Mapping examines the disagreements that persist. Meta-Analysis brings out the implicit structures shaping the positions. Critical Archaeology traces the formulations, vocabularies, exclusions and framings that delimited the discussion. Horizon of Possibilities formulates what the debate made thinkable without stating it. Argumentative Evaluation and Source Verification test, respectively, the quality of the argumentation and the strength of its factual support.

The ideas carried by Syntia and Dyfonia did not disappear. Their boundaries did.

Transformation as an effect of trajectory

In practice, the architecture revealed something the original triad represented poorly: the transitions between modes matter as much as the modes themselves. A convergence reached after a critical detour does not have the same history as one that was never challenged. Its formulations have met objection: some were abandoned, others reformulated or strengthened.

The detour is not a test of truth. A true thesis may give way to a brilliant objection; a weak position may survive through skilful defence. But a formulation that has come through challenge no longer has the same argumentative status.

A Hegelian intuition illuminates this point. Negation is determinate when its result retains the content it negates while bringing forth a new determination. The analogy has its limits; it at least explains why confrontation can be productive without opposition amounting to proof.

Metamorphosis also operates at a second level. An analysis remains external to the debate it describes: naming a presupposition or a persistent tension does not alter the exchange after the fact. It becomes operative when it is fed back into the debate.

Reused verbatim, the follow-up question produced by certain analysis modes becomes a new object of debate. Tension Mapping can bring the models back to the precise point where their criteria of judgement diverge. Emergence Analysis can return them to a concept they have just built and ask them to test or extend it. Meta-Analysis can put a shared axiom or blind spot back before them.

Analysis then becomes a transition operator, and the process becomes a loop: the debate produces its material; analysis rereads it from a particular angle; that reading generates an intervention; the intervention changes the material available to the exchanges that follow.

This is where the name finds its meaning. Metamorphosis is not the operation of any one mode. It unfolds along a trajectory: changes of mode, reformulations prompted by objection, concepts born from friction between positions, the rereading of the whole through analysis — which sends back a question the debate had not yet asked.

Why the name remained

Within this architecture, Syntia and Dyfonia no longer had modules to name. Their logic remains: Syntia survives whenever several positions work out a connection that none of them yet held in that form; Dyfonia whenever a disagreement lasts long enough to force the positions to make their differences explicit.

Metamorfon had a different fate. Initially the third term of the triad, it came to name the product of the other two moves — and later of those that followed — when combined over time. The name of the application became the name of its intended effect, rather than of any particular means.

Convergence builds; dissonance tests. In the passage from one to the other, sometimes, something is transformed. The name preserved that possibility.

References

• H. G. Liddell, R. Scott and H. S. Jones, A Greek-English Lexicon — entries σύν, διαφωνία, φωνή, μεταμόρφωσις and μεταμορφόω, Scaife ATLAS / Perseus Digital Library.
• Douglas Harper, Online Etymology Dictionary — entry ‘dys-’.
• G. W. F. Hegel, The Science of Logic, trans. George di Giovanni, Cambridge, Cambridge University Press, 2010, Introduction — on determinate negation and the way its result retains the content it negates while producing a new determination.

L’article Why Metamorfon? est apparu en premier sur Metamorfon.

On the models’ side, adaptation is indeed nearly automatic: each model adjusts its contribution to the debate mode in effect. A critical mode pushes it to expose the flaws of the opposing position, a constructive mode to seek possible articulations, a balanced mode to hold both at once. This kind of adaptation requires nothing from the user — it occurs simply by virtue of the turn’s settings.

On the user’s side, by contrast, adaptation is anything but automatic. It demands an active reading. At each turn, the user observes what the models have just produced — where they converge, where they resist, what they have avoided, what calls to be tested — and chooses the next mode based on what is observed and on the objective being pursued in that session. A debate that gets bogged down in a terminological quarrel will benefit from a shift to the constructive mode to force an articulation; a debate that converges too quickly calls for a shift to the critical mode to put that convergence to the test; a debate that produces many arguments but little movement is best served by a shift to the refutational mode to settle the matter. No rule dictates these choices; they fall to the user’s judgment about the state of the debate and about what is sought from it.

It is the articulation of these two adaptations — that of the models, which follows the mode, and that of the user, who chooses the mode — that produces the trajectory of a session. The adjective adaptive names precisely this articulation. It designates neither a hidden automatism nor an intelligence of the device that would dispense with human steering. On the contrary, it designates the condition of possibility of this steering: a device where each turn opens a choice, and where that choice belongs to the one who conducts the session.

This is also why the experience of using Metamorfon refines itself with practice. Early uses fumble; over the course of sessions, the gesture becomes more apt — the user more quickly recognizes what calls for a shift, anticipates what a mode will produce on the current state of the debate, composes trajectories that no isolated choice could have built. It is in this accumulated gesture, more than in any feature taken in isolation, that the know-how specific to the use of the tool resides.

L’article What adaptive means est apparu en premier sur Metamorfon.