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.

In any discussion, part of the work has already been done before anyone responds. The wording of the initial question has selected certain terms, installed certain distinctions, identified some objects as relevant, and left other possibilities at the periphery.

Consider a question such as: “Granting that something is shifting in the creative gesture today — in the arts, but also wherever the task is to produce the new — what should we call what is shifting? Not the instrument that may be causing it, but what is being transformed in the act of creating.”

It sounds open. Yet its wording already distributes the space of possible answers. “Granting that something is shifting” foregrounds transformation rather than continuity. “Not the instrument” brackets one level of explanation in order to direct attention towards the act of making itself. And asking what the shift should be “called” favours answers capable of naming and conceptualising a transformation.

None of this makes another answer impossible. A model can challenge the premise, reject the distinction between instrument and act, or refuse the demand for a new concept. But doing so requires it to step partly outside the frame it has been given.

This is the layer that Critical Archaeology seeks to make visible: not what the debate literally could not have said, but how its framing made some formulations immediately available, others less likely, and some alternatives dependent on a prior displacement of the question itself.

Foucault’s Archaeological Gesture

The principal theoretical reference point for the mode is Foucauldian archaeology.

In The Archaeology of Knowledge, Michel Foucault does not search behind discourse for a hidden intention or for a deeper meaning unknown to the speakers themselves. He examines the regularities that organise, within a given discursive field, the formation of objects, concepts, enunciative positions, and strategies.

The task is therefore not only to ask whether a statement is true or false, well or poorly argued, but to shift attention towards the conditions under which certain types of statement become available and intelligible.

Critical Archaeology does not apply this method in the strict Foucauldian sense to a conversation between AI models. It retains a methodological gesture from it: not to examine only the answers that were produced, but to interrogate the field within which they took this form — the terms of the question, the distinctions it establishes, the categories it privileges, and the alternatives it leaves at the margins.

A deterministic reading should be avoided here. A discursive frame does not operate like a mechanical boundary that permits certain sentences and absolutely forbids others. It organises a distribution instead: some positions are immediately available within it, some are marginalised by its formulation, and others would require a displacement of the very categories in which the debate was opened.

Framing as a Distribution of Possibilities

This inquiry begins with the user’s question. Every question makes choices. It designates an object, selects a vocabulary, sometimes implicitly assigns value to particular states of affairs, and indicates what kind of answer appears relevant.

Those choices are not necessarily mistakes. A question entirely devoid of framing would probably be unusable: asking something already requires dividing up a problem in a particular way. The purpose of Critical Archaeology is therefore not to imagine a “neutral” question that could escape every frame. It is to make the chosen frame visible enough for that frame to become an object of examination in turn.

The same applies to later user interventions. A follow-up can open a path that had remained marginal, but it can also narrow alternatives by demanding a number, imposing a time horizon, asking the models to choose between two options, or reformulating the problem through a new distinction.

The mode therefore analyses interventions not only for their argumentative content, but as acts of framing that redistribute the space within which the conversation can continue.

Reading a Question as a Structure

The session What is shifting in the creative gesture today? provides a concrete example of this mechanism.

The discussion produced several distinct concepts and eventually a common framework for describing transformations in creative agency. Yet several features of that architecture had already been prepared by the initial wording of the question.

The phrase “granting that something is shifting” first creates an asymmetry. It makes accounts of transformation particularly available, whereas the possibility that nothing fundamental has changed requires an explicit challenge to the premise.

This does not prevent a model from arguing that the apparent transformation is superficial, institutional, commercial, or merely technological. But such an answer must first reopen something the question has provisionally asked the debate to hold constant.

The instruction “not the instrument that may be causing it, but what is being transformed in the act of creating” introduces a second division. It foregrounds the creative act and deliberately backgrounds explanations centred on infrastructure, ownership, training regimes, labour, or political economy.

That division is methodologically productive: it allows the session to concentrate on a specific level of analysis. But it also has an analytical cost. Once the instrument has been bracketed, accounts in which the transformation of creative practice cannot be separated from its technical and economic infrastructure become less immediately available.

Finally, asking “what should we call what is shifting?” favours a particular form of response: conceptual naming. The models duly propose a poietic regime, a displaced threshold, and ontological steering, before progressively organising them into a common architecture.

Again, the question does not force them to invent concepts. But conceptual coinage fits the task particularly well, while a response such as “the phenomenon does not require a new name” would have to contest the requested operation itself.

The answers therefore adopt much of the architecture made salient by the question. This is consistent with the framing, but it does not establish that the framing is the sole cause. The models also bring their own regularities, training corpora, conceptual repertoires, and response habits.

The archaeological interest lies precisely in maintaining that distinction: not explaining every answer causally through the prompt, but showing how strongly the answers and the initial frame already share a grammar of the problem.

The vocabulary developed by the models then reinforces particular directions. Poiesis, authorship, agency, threshold, possibility space, individuation, selection, constraint: the debate increasingly organises creation as a problem of where and how agency operates within a structured process of making.

That vocabulary makes some questions particularly accessible — where authorship sits, where commitment becomes irreversible, how a possibility space is configured. It leaves other questions further towards the margins: who owns the infrastructure, whose labour sustains it, how markets define the value of “creation,” or whether the category of authorship remains the right organising principle at all.

The session itself eventually encounters part of this limit. Its later turns ask whether art still requires a human author and whether its inherited vocabulary of poiesis continues to name the phenomenon adequately. The subsequent editorial and verification work also brings labour, infrastructure, and closure back into view.

This does not mean that the original question was poorly framed. Critical Archaeology makes visible the analytical cost of its framing: the possibilities it renders particularly accessible and those that require an additional move to bring back into the field.

From Archaeology to Genealogy

A distinction is needed here. Making visible the rules and categories that organise a discourse is not exactly the same operation as tracing their history.

In Foucault, archaeology and genealogy are related, but they are not interchangeable. Archaeology is concerned in particular with describing discursive formations and their rules. Genealogy places greater emphasis on descent, emergence, discontinuity, and the contingent historical processes through which particular practices and categories have taken shape.

This second orientation is developed explicitly in “Nietzsche, Genealogy, History”, first published in 1971.

Metamorfon’s Critical Archaeology can draw on both gestures, but they should remain distinct.

When it shows that a question rests on a distinction between instrument and act, a vocabulary of authorship, or a particular understanding of creativity, its work is primarily archaeological: it describes the categories and regularities organising the field of the debate.

When it goes further and asks how those categories were historically constituted, which traditions stabilised them, or which transformations gave them their present form, the gesture also becomes genealogical.

Keeping this distinction clear prevents every historical reference from being treated as “archaeology” and makes more precise what the analysis is claiming to establish.

Distinctions From the Other Modes

Critical Archaeology differs first from Meta-Analysis.

Meta-Analysis examines what structures the debate within the frame it has been given: admitted axioms, epistemic styles, blind spots, and patterns of convergence or divergence. Critical Archaeology shifts the analysis towards the frame itself: why does this problem appear in this form? Which categories has the question already mobilised? Which alternatives would require those categories to be displaced?

An image remains useful. Meta-Analysis observes a chess game and examines the players’ strategies, preferences, and blind spots. Critical Archaeology also asks why the chessboard has this form and which rules defined the moves as relevant.

The mode is not critical in the evaluative sense either. Bringing a framing to light does not mean that the framing is bad. A formulation can be entirely legitimate while still privileging particular perspectives.

The operation is first to make the frame visible and contestable as a frame, where it might otherwise have appeared self-evident. Once visible, the user can retain it, modify it, or test another.

Critical Archaeology is also not a hermeneutics of suspicion. It does not search for hidden intentions in the interlocutors or presume that a concealed interest lies behind their arguments. It examines observable discursive conditions: vocabulary, oppositions, presuppositions, and the architecture of questions.

A debate can be perfectly sincere while operating within a frame that no participant explicitly selected.

Nor is the mode an exercise in systematic deconstruction. Its purpose is not to dissolve the categories it identifies or to demonstrate their illegitimacy.

It places them within a particular field and shows that they are not the only possible way of organising the problem. When it also traces their historical descent, it adds a genealogical dimension; that still does not imply that the categories should be abandoned.

Convergence Under Shared Constraints

This mode becomes particularly interesting when several models converge.

Convergence may of course result from arguments that each model finds convincing. But it may also depend in part on elements that were already shared before the confrontation began: vocabulary supplied by the question, an opposition that nobody challenged, a common normative horizon, or alternatives left outside the frame.

Critical Archaeology should not decide in advance that every consensus is artificial. It can, however, test the hypothesis of a shared constraint.

If three models arrive at a similar recommendation while retaining different arguments, it becomes useful to ask which elements of their discursive environment remained constant. Did they genuinely converge on the justification? Did they chiefly accept the same definition of the problem? Do certain solutions appear common because the question made them particularly available from the outset?

The creative-gesture session offers a useful example. Its three initial concepts differ substantially, yet they are rapidly reorganised as complementary levels of a single framework. That convergence may reflect genuine argumentative accommodation. At the same time, all three models were responding to a question that had already foregrounded transformation, bracketed the instrument, and requested conceptual naming. The convergence can therefore be examined simultaneously as an intellectual achievement and as something formed within shared upstream constraints.

This distinction turns agreement into an object of analysis without disqualifying it.

A consensus can be robust and depend on a shared frame. The two properties are not mutually exclusive.

Uses of Critical Archaeology

The mode is particularly useful when a debate is intended to contribute to an important decision. A strategic choice, an editorial direction, or an institutional position often depends on a question formulated upstream. Identifying what that question privileged helps separate the answer obtained from the frame within which it was obtained.

It is also relevant when convergence appears unusually rapid. When several models seem spontaneously to reach similar conclusions, analysing the frame can help determine whether their agreement concerns the arguments themselves or also rests on shared presuppositions.

In research, journalism, essay writing, or consulting, it can be used to test the formulation of the problem before building further reasoning upon it. It does not necessarily produce a better question; it makes the choices already embedded in the existing question more explicit.

The mode is especially productive when the subject of the debate concerns the conditions of discourse themselves: language, information, political communication, the production of categories, or the authority to speak. In such cases, the debate can be subjected to the same questions it applies to its own object.

For more direct purposes, other modes remain better suited. If the aim is to reconstruct the course of the session, Integrative Synthesis is sufficient. If the aim is to examine its axioms or epistemic styles, Meta-Analysis is more direct. If the aim is to identify what the progression of the debate has now made formulable without itself formulating it, Horizon of Possibilities is the relevant mode.

The value of Critical Archaeology comes precisely from this specialisation: it shifts attention from positions towards the conditions that helped give those positions this particular form.

A Methodological Limit: The Analysis Window

Like Metamorfon’s other analysis modes, Critical Archaeology works directly from the ten most recent turns in their full verbatim form; earlier turns, when they remain necessary to the context, are provided in summarised form.

This architecture requires particular caution for this mode.

In a long session, the opening question and the earliest interventions may no longer belong to the raw window actually submitted for analysis. Yet Critical Archaeology may specifically want to examine the choice of a word, a grammatical opposition, an alternative implicitly opened by the initial formulation, or a shift introduced several turns earlier.

Fine-grained lexical analysis therefore does not have the same degree of robustness when the relevant formulation is directly available in the verbatim as when it is known only through an earlier summary.

Summarised context may be sufficient to identify a general orientation — for example, that the session opened within a normative or economic frame — without supporting with equal precision the claim that a particular word shaped the architecture of the exchange.

The scope of the conclusions must therefore be read at the scale of the material actually available: the more an interpretation depends on a precise lexical detail, the more it requires direct access to the formulation from which the argument is drawn.

This limitation does not make the mode unusable in long sessions; it simply distinguishes what it can establish at fine resolution from what it can reconstruct only at a more general level.

Making the Frame Contestable

Critical Archaeology ultimately searches neither for a hidden meaning behind the answers nor for a single cause explaining why the models spoke as they did.

It works on a closer and more observable layer: the way a question and the progression of an exchange distributed the objects, categories, and directions available to the debate.

Its gesture is not to declare those choices wrong. It is to move them from the status of self-evidence to that of examined choices.

An invisible frame can easily present itself as the natural form of the problem. Once made visible, that frame can be retained, challenged, compared with another, or reformulated deliberately.

This is the sense in which the mode is critical: not because it refutes the frame, but because it prevents that frame from remaining entirely exempt from the examination it imposes on the answers.

It moves from the answers that were produced back towards the conditions that oriented the form in which they could appear — not in order to recover the hidden cause of the debate, but to make visible what its frame was already doing before the positions began to confront one another.

References

• Michel Foucault, The Archaeology of Knowledge, translated by A. M. Sheridan Smith, Routledge — discursive formations, rules of formation, and the conditions under which statements appear and are transformed.
• Michel Foucault, “Nietzsche, Genealogy, History,” in Language, Counter-Memory, Practice: Selected Essays and Interviews, translated by Donald F. Bouchard and Sherry Simon — descent, emergence, discontinuity, and the genealogical gesture.

L’article Critical Archaeology est apparu en premier sur Metamorfon.

At the end of an intense conversation, one sometimes has the sense that “we touched on something important without ever formulating it.” The exchanges seem to have prepared an idea, a distinction, or a question that no one ultimately stated. Something remained on the horizon — perceptible from what was said, but never explored.

This experience points to a precise phenomenon: a debate can produce more than its explicit statements. By bringing certain positions into contact, exposing tensions, exhausting particular frames, or progressively shifting a question, it can give new formulations an argumentative grounding they did not have at the outset. It establishes a conceptual space of which it explores only part. It prepares questions it does not ask.

The Horizon of Possibilities is the analysis mode designed to examine this layer. Not what was said — that is the work of Integrative Synthesis —, not what was actually constructed and formulated through the exchange — that is the work of Emergence Analysis —, but what the debate made formulable without formulating it.

A Deleuzian Echo

The gesture of the Horizon of Possibilities finds a particularly illuminating reference point in Gilles Deleuze’s distinction between the possible and the virtual.

In Difference and Repetition, Deleuze insists that the virtual is opposed not to the real but to the actual: it possesses a reality of its own and is actualised through differentiation, whereas the possible belongs to the possible/real pair and is realised.

This distinction also extends a critique developed by Henri Bergson. In “The Possible and the Real,” collected in The Creative Mind, Bergson challenges the idea that possibilities exist fully formed before their realisation. The possible is often projected retrospectively behind what has actually occurred: novelty is not simply a matter of selecting among forms that were already available.

The Horizon of Possibilities does not apply this ontology to a conversation. It retains an operational distinction from it: one should not confuse a list of continuations that could freely be imagined with directions whose formulation has been prepared by the actual structure of the debate. An unresolved tension, an implication left undrawn, or an unthematised shift of frame can make certain formulations better grounded at the end of a session than they were at its beginning.

The term “virtuality” therefore functions here as a methodological analogy, not as a philosophical identity. For Metamorfon, a virtuality is admissible only if the analyser can show from which traces in the debate it has been constructed. This requirement of grounding is what distinguishes the Horizon of Possibilities from an exercise in free speculation.

Formulating What the Debate Has Prepared

Concretely, the mode performs a particular operation. It rereads the session and formulates continuations whose grounding in what actually occurred must be made visible.

The question is not what the models might have said under different instructions, with more knowledge, or with greater imaginative freedom. That would open an almost unlimited space of speculation. The Horizon works instead from constraints supplied by the corpus: conceptual beginnings, unresolved tensions, undrawn implications, progressively displaced frames, and questions left open.

For this operation to remain rigorous, every identified horizon must be connectable to actual passages in the debate. The mode does not claim to discover a hidden idea that somehow pre-existed within the conversation; it constructs a formulation from elements whose combination makes that formulation defensible.

The distinction is crucial. It separates a grounded inference from a creative production that merely uses the session as a pretext.

The mode can in particular reveal three types of horizon.

First Type: Concepts One Word Away

Several elements in a debate may converge towards a distinction that no one explicitly formulates. The interlocutors possess the components of the concept and may even oppose them to one another, yet no voice gives them a common form.

In a session on whether AI systems should be allowed to refuse instructions on ethical grounds, one such possibility formed around irreversibility. The debate had progressively weakened “harm” as a stable foundation for refusal and introduced reversibility and irreversibility as conditions for deciding when an ex-ante refusal might still be justified.

The seventh analysis drew out a conceptual implication that the debate itself had not developed: time, rather than harm, could become the operational foundation of refusal. A refusal architecture indexed to reversibility would be a fundamentally different object from one indexed to harm.

The elements were present in the session, but their combination into that proposition was not. This is precisely the kind of formulation the Horizon of Possibilities must present as inferred from the corpus, rather than discovered in it.

Second Type: Structural Shifts in the Problem

A debate may also progressively shift its own object while addressing its initial question. The interlocutors begin by examining one problem, but their answers and objections bring another level of questioning into view without explicitly thematising the substitution.

The same session began as a debate about whether AI systems should be permitted to refuse instructions on ethical grounds. Yet as the exchange developed, the models increasingly discussed standing, burden of proof, appeal, review procedures, institutional oversight, and the conditions under which decisions could acquire legitimacy.

The seventh analysis identified the displacement: the conversation had quietly moved from moral philosophy towards something resembling constitutional design. The issue was no longer primarily whether an AI refusal was ethically correct, but what kind of institutional order could make such a decision legitimate.

No participant explicitly announced this change of object. The Horizon of Possibilities reconstructed it from the trajectory of the session.

Identifying this kind of displacement has practical value: it can reveal that the question actually being worked on at the end of a session is no longer exactly the one that was asked at the beginning.

Third Type: Questions the Debate Prepares Without Asking

After several turns, some questions acquire a grounding they did not possess at the start of the session. The exchanges have explored certain answers, encountered their limits, introduced distinctions, or shifted the frame; a new question therefore becomes better grounded.

In the same debate, one formulation became central: the legitimacy of a refusal could be earned retrospectively by surviving contestation rather than secured in advance by claiming that the refusal was substantively correct.

The seventh analysis then opened a question the debate itself had not pursued: if refusals require post-hoc legitimation, why should compliance not require it as well? If an AI’s refusal is provisional until it survives challenge, perhaps an AI’s decision to comply has the same provisional structure.

At Turn 0, that question would not have had the same argumentative grounding. It depends on the session’s prior movement from substantive harm prevention to procedural legitimacy and retrospective justification.

Nor is it a summary of what the models said. It extends a principle they had developed into a domain they had left asymmetrical. This is precisely how the Horizon of Possibilities can formulate a new question without treating it as though it had already been present in the corpus.

What Distinguishes the Horizon of Possibilities From Other Modes

A first distinction is essential because it concerns two very different ways of speaking about novelty. The Horizon of Possibilities is not Emergence Analysis.

Emergence Analysis concerns what is attested in the corpus: a formulation, distinction, or concept that actually appeared during the exchange without having been present in that form at its starting point. Such elements can be quoted, located in a specific turn, and followed through their subsequent uptake and transformation.

“Legitimacy debt” in a session on citizen consultations, “revocable infrastructure” in a session on machine speech, “axiological fork” in a session on AI ethics: when such formulations actually occur in the responses, Emergence Analysis can document their appearance and trace their trajectory.

The Horizon of Possibilities works with a different status. What it formulates is precisely not attested as such in the corpus. It must be inferred from it: an unasked question, an undrawn consequence, a distinction that several passages make possible to construct.

When the Horizon formulates “democratic sovereignty is shifting towards material layers” or proposes “learning to deliberate without the human/machine distinction,” those statements appear in no turn. The task is precisely to show why their formulation becomes defensible from the conjunction of elements that are present.

The two modes therefore complement one another because they work on two different statuses of novelty. Emergence Analysis concerns what was actually formulated; the Horizon of Possibilities concerns what must still be inferred from the material.

The Deleuzian distinction between the actual and the virtual provides an illuminating analogy here, but it should not be treated as a strict equivalence. Within Metamorfon, the operational boundary is simpler: on one side, what is attested; on the other, what the analysis can reasonably construct from the traces left by the debate.

A second distinction concerns Tension Mapping.

Tension Mapping analyses what continues to resist after confrontation: persistent disagreements, their nature, their depth, and the conditions under which they might eventually shift. The Horizon of Possibilities does not primarily ask what resists, but what this configuration still opens.

The two can intersect. A persistent tension may become the point of departure for a question no one has yet asked; a disagreement over the criteria of judgement may make necessary a shift of level that the interlocutors themselves never undertook.

But the analytical gesture remains different: Tension Mapping describes the structure of what remains in tension; the Horizon constructs, from that structure and other traces in the debate, directions that remain to be explored.

When to Use the Horizon of Possibilities — and When Not To

The Horizon of Possibilities is a particularly demanding mode because its object is not directly present in the form in which the analysis reconstructs it. It is therefore neither the first mode to activate on every session nor one to use indiscriminately.

Situations Where It Is Particularly Relevant

When a session is complete and one wants to derive directions for further work from it. A session does not necessarily exhaust its object. The Horizon of Possibilities can identify questions, distinctions, or shifts that the session has made better grounded without directly exploring them. It can therefore point towards a new session, an article, or a subsequent line of reflection.

For anyone using Metamorfon as a research or creative tool, it can play a generative role.

When looking for new questions on a subject that already seems heavily explored. Some topics appear to have been approached from every angle — AI ethics, deliberative democracy, algorithmic governance. By starting not from the topic in general but from the precise trajectory of a particular debate, the Horizon can reveal situated questions that the usual categories of the field did not bring into view.

When one wants to avoid stopping too early at an apparent consensus. Convergence can leave implications unexplored. The Horizon can then ask what the agreement has had to leave undetermined, or what would be put under pressure if its reasoning were pursued further.

When preparing editorial or intellectual work. For a researcher, journalist, essayist, or consultant working from a session, the mode can identify angles that have not yet been addressed but are sufficiently grounded in the material to deserve development.

When a session has produced a relatively stabilised object — a protocol, specification, or architecture — and one wants to examine what that stabilisation has left aside. Converging on a particular construction often means excluding alternatives, fixing a vocabulary, and suspending certain questions.

The Horizon of Possibilities reverses the perspective: rather than reading the session only for what it constructed, it examines the directions that construction did not retain and asks which of them remain relevant in light of the elements actually present.

In that respect, it is a particularly useful complement to an Emergence Analysis of an architectural session: one describes what was constructed through the exchange; the other examines what that construction still leaves open.

Situations Where the Horizon of Possibilities Is Less Appropriate

When one simply wants to understand what was said. An Integrative Synthesis is more than sufficient. The Horizon already presupposes a solid understanding of the content; read in isolation, it risks blurring the distinction between what belongs to the debate itself and what belongs to the analyser’s inference.

When the debate is very short or conceptually underdeveloped. The mode needs sufficiently rich material for its inferences to be constrained by several converging elements. Applied to only a few turns that have barely shifted the positions, it risks producing generic continuations that almost any discussion on the same subject could have suggested.

When one needs a strictly factual and literally traceable reconstruction. By design, the Horizon formulates statements that do not appear verbatim in the debate. They must be justifiable from what was said, but they cannot be treated as quotations or as conclusions directly stated by the participants. In contexts where that distinction is critical — a strict methodological audit, or some legal or documentary analyses — other modes are more appropriate.

The Mode’s Specific Risk: Overinterpretation

This feature exposes the Horizon of Possibilities to a greater methodological risk than modes that operate mainly on directly attested elements: overinterpretation.

Emergence Analysis can point to the formulation whose appearance it traces. Tension Mapping can quote the positions it places in relation. Integrative Synthesis can refer back to the stages it reconstructs. The Horizon, by contrast, formulates precisely something that does not appear as such in the corpus.

Its quality therefore depends on its ability to make its inference auditable.

Every proposed horizon should be connectable to identifiable elements in the session — a tension, an undrawn implication, a shift of frame, a contradiction, an unanswered question — and should make visible the path leading from those traces to the proposed formulation.

The shorter that path is, and the more strongly it is constrained by several converging elements, the more firmly grounded the Horizon is. The more it requires intermediate assumptions, generalisations, or elements absent from the corpus, the more cautiously its status should be presented.

This requirement also makes it possible to distinguish two different judgements: a formulation can be intellectually interesting without being a well-grounded Horizon of the session. The mode succeeds only when both conditions are met: it opens something productive and makes visible why that opening arises from the material analysed.

Like Metamorfon’s other analysis modes, the Horizon of Possibilities works directly from the ten most recent turns in their full verbatim form; earlier turns, when they remain necessary to the context, are provided in summarised form.

Its most precise inferences should therefore be read first at the scale of this raw analysis window. Earlier context may help illuminate them, but a virtuality, tension, or implication that the analyser can recover only from an older summary does not have the same degree of traceability as an element directly present in the ten turns being analysed.

What the Horizon of Possibilities Adds Beyond Its Results

This analysis mode does more than produce a few possible continuations at the end of a session. It also proposes a particular way of rereading what has just taken place.

Used regularly, the mode can encourage greater attention to undrawn implications, to shifts in the question that do not announce themselves as such, to concepts whose components appear without ever being assembled, and to the problems that an agreement still leaves open.

This way of reading does not consist in assuming that a hidden idea lies behind every sentence. It requires the opposite discipline: distinguishing what is present, what is implied, what becomes plausible through combination, and what would merely be a projection by the analyser.

That may be the most interesting function of the mode. It does not simply ask: “what more could have been said?” It asks:

what does this particular session now give us good reasons to formulate that we would not have had the same reasons to formulate before it?

That question keeps the Horizon of Possibilities in its proper place. It is neither a prediction of what the debate “contained” nor the free generation of new ideas, but work on the gap between what was actually said and what the progression of the dialogue now gives us stronger reasons to think.

References

• Gilles Deleuze, Difference and Repetition, translated by Paul Patton, Columbia University Press — the distinction between the possible and the virtual, the reality of the virtual, and the process of actualisation.
• Henri Bergson, “The Possible and the Real,” in The Creative Mind: An Introduction to Metaphysics, translated by Mabelle L. Andison — critique of the idea that fully constituted possibilities pre-exist the real, and reflection on the creation of novelty.

L’article Horizon of Possibilities est apparu en premier sur Metamorfon.

Among Metamorfon’s eight analysis modes, Integrative Synthesis is at once the most immediately useful and the least spectacular. The most immediately useful, because anyone who has just launched a nine-turn debate between two or three models has a simple first need: to know what was said.

The least spectacular, because synthesizing content does not appear to involve a distinctive cognitive operation — it is, one might think, what every model does all day long.

This apparent simplicity conceals the difficulty specific to the mode. Synthesizing a debate is not the same as summarising a text. A text comes with a given organisation — an author, a progression, a thesis — and the summary follows what the author has laid out. A debate, by contrast, contains several voices, sometimes non-synchronous, that take up, reformulate, contest, and appropriate one another’s arguments, and whose argumentative trajectory is not linear.

This specific difficulty of dialogue summarisation is also documented in natural language processing research: coreference, dialogue acts, and interactions between speakers introduce challenges that do not arise in the same way when summarising continuous written text (Xu, Zhu & Zeng, EMNLP 2022).

What one model said at Turn 3 may be reformulated at Turn 5, refuted at Turn 6, and reformulated differently at Turn 8 after a user intervention. Reconstructing this movement — who said what, when, in response to what, and with what effect — requires something more than compression: it requires a narrative reconstruction of the debate’s conceptual trajectory.

Narrative theory has a name for this operation: emplotment.

Emplotment According to Ricœur

In Time and Narrative, Volume 1, Paul Ricœur develops a concept that offers a particularly useful way of understanding what an Integrative Synthesis does to a debate. He calls emplotment — taking up and transforming the Aristotelian mythos — the operation through which a succession of heterogeneous events is gathered into an intelligible story.

For Ricœur, emplotment does not simply add a narrative to facts that are already there. It does two things at once: it selects from among events those that matter to the trajectory being traced, and it configures those events by relating them within an intelligible whole. A chronicler can list what happened; a narrator turns it into a story. Emplotment transforms multiplicity into sequence — a succession of things into a progression that makes sense.

This operation is more than an editorial convenience. It responds to what Ricœur calls the aporias of lived time: time in its pure flow resists understanding; time configured through narrative becomes thinkable. A debate presents a comparable problem: the raw sequence of turns is full of repetitions, reformulations, and tacit reprises that resist immediate legibility.

Without emplotment, a reader returning to a long debate can lose the thread — each turn appears equivalent to every other, without making clear why one position was refuted, why a particular concession mattered, or why an intervention redirected the trajectory.

Integrative Synthesis does not apply Ricœur — no more than the other modes apply Foucault, Deleuze, Whitehead, Simondon, or Lyotard. But its fundamental gesture — recomposing a succession of turns into an intelligible argumentative trajectory — finds in the notion of emplotment a particularly illuminating reference point. Its purpose is not to add positions that do not appear in the debate; it selects and reconfigures what actually occurred in order to make its progression legible.

What the Synthesis Does — and What It Does Not Do

A well-constructed Integrative Synthesis reconstructs four interlocking layers of a debate, in an order that is not negotiable.

It first reconstructs the initial framing — the positions set out by each model in the opening turn, with their emphases, organising categories, and visible presuppositions. This is the raw material of the narrative that follows, and a poorly reconstructed initial framing distorts everything that comes after.

It then reconstructs the argumentative trajectory — who introduced a frame, concept, or criterion; what the other voices did with it; which objections were raised; which positions were reformulated; which concessions were made. An element may enter through a user intervention: the narrative attributes it to that intervention rather than crediting it to a model. This is the heart of emplotment: what turns a succession of turns into a coherent movement.

The narrative follows the circulation of elements, not a ranking of participants.

This reconstruction is not a simple sequential listing. It explicitly identifies the pivots of the debate — moments at which its orientation, tone, or frame shifted — and classifies them according to a precise typology: concession (a model accepts a point it had contested), reversal (a model changes its position), frame shift (the conceptual reference point moves without the models explicitly thematising the change), and structuring user intervention (a question or follow-up that redirects the debate).

A withdrawal of an argument is therefore classified as a concession; another session, devoted to the very notion of intelligence in AI, sees its trajectory pivot around a structuring user intervention that deliberately narrows the technical framing, followed by a frame shift that the models themselves do not explicitly identify.

For each pivot, the synthesis does not merely state that it occurred. It explains why it functioned as a pivot — what it made possible, what it displaced, and what subsequent turns did with it. It is this requirement of trajectory reconstruction that turns synthesis into analytical work rather than a simple report.

And the same discipline applies negatively: when a debate contains no genuine pivot, the synthesis states this explicitly rather than inventing one to fill the category.

For every pivot, the narrative also indicates which voice introduced the element that shifted and what the other voices subsequently did with it. This attribution is anchored to a dated trace — a turn, a formulation. It describes what a voice does at a given moment, never what that voice is.

Before attributing an uptake, the narrative identifies the architecture of the session. In a crossed debate or a trilogue, the voices within a single turn respond simultaneously, without having read one another: a resemblance within the same turn is not an uptake. Circulation from one voice to another can only be established from the following turn onward. In an alternating dialogue, the second voice has read the first: a resemblance may be an uptake, but similarity alone does not establish it.

The synthesis reconstructs, third, the conceptual shifts and developments — moments when a position changed, a concept was enriched, or one frame was replaced by another.

This is where Integrative Synthesis overlaps — without duplicating — Emergence Analysis: it mentions co-constructed concepts such as differential friction, scale neutrality, or the irreducible residue because they are part of what was said, but it does not specifically map them across the seven axes of the emergence mode. The Synthesis incorporates them into the narrative without making them the object of a separate examination.

Finally, it reconstructs the points of arrival — what the models converge on at the end of the debate, what remains disputed, and what remains undetermined for lack of information. This overlaps with Tension Mapping, but without the same analytical depth: it does not classify disagreements according to their nature and resolvability; it records them as part of the narrative outcome.

What Integrative Synthesis does not do is just as important as what it does. It does not judge the argumentative quality of the positions — an argument may be poorly grounded and still be faithfully reported as a moment in the debate. It does not deconstruct implicit framings — that is the work of Meta-Analysis and Critical Archaeology. It does not project towards what the debate might have produced but did not — that is the work of the Horizon of Possibilities.

It remains within the register of faithful reconstruction, while accepting that fidelity requires intelligible reconfiguration rather than complete transcription.

A Stable Structure, Multiple Styles of Execution

Integrative Synthesis follows a stable structure: the narrative of the debate, followed by a question addressed to the models. This structure remains constant, but the grain and register of execution can vary substantially depending on the analysis model, its instructions, and the material being analysed.

One style privileges continuous narrative. The synthesis follows the succession of turns in fluid prose, reconstructing the principal arguments, concessions, and shifts. The reader can therefore follow the debate almost as a story, but in a condensed form organised around its significant pivots.

A second style privileges interpretive condensation. The narrative is shorter and seeks to bring a small number of memorable formulations or moments to the foreground. It sacrifices some of the granularity of the verbatim but makes the shifts that structure the overall trajectory more salient.

A third style privileges explicit structure: numbered sections, categories, lists, or tables that make it possible to locate positions, agreements, disagreements, and interventions quickly. It sacrifices some narrative continuity in exchange for navigability.

In the examples compared on Metamorfon, these forms of reconstruction generally recover a common substantive core — principal pivots, structuring interventions, residual disagreements — while distributing narrative weight differently. The choice of analyser can nevertheless also affect what is retained, grouped together, or left in the background. These differences matter less as “styles” to choose between than as a methodological reminder: a synthesis is always a selective reconstruction of its material, and its fidelity is therefore measured not by uniformity of form but by the accuracy of its attributions, connections, and reconstruction of the trajectory.

A Particular Requirement: Reliability of Reconstruction

Among Metamorfon’s eight analysis modes, Integrative Synthesis is the one in which fidelity to the material carries the greatest weight. The other modes produce analyses whose value depends in part on the framework they apply; a synthesis, by contrast, has reliability as its primary requirement. If it distorts what was said, it no longer correctly reconstructs the trajectory it claims to describe.

This requirement has consequences for the choice of analysis model. Dialogue summarisation is particularly sensitive to speaker tracking: research has shown that even powerful summarisation models can fail to distinguish speakers correctly and thereby introduce factual attribution errors (Geng et al., COLING 2022).

In Metamorfon, an error of this kind — attributing to Model B an argument introduced by Model A — is particularly serious, because it alters precisely the trajectory that the Synthesis is supposed to reconstruct.

Another attribution error, specific to crossed debates, consists in interpreting as an uptake what is only a within-turn resemblance: because the voices have not yet read the other responses from the same turn, circulation cannot be established there. The synthesis must therefore track not only speakers but the exact temporal architecture of the session.

This requirement becomes more important as the material grows longer. Having access to a long context window does not guarantee that all of the information it contains will be used with equal reliability: model performance can vary, in particular, according to where the relevant information occurs within the context (Liu et al., TACL 2024).

In Metamorfon, analysis modes work directly from the ten most recent turns in their full verbatim form; earlier turns, when they remain necessary to the context, are provided in summarised form. Integrative Synthesis can therefore take that earlier context into account, but its finest-grained attributions and detailed reconstruction of the trajectory rest first on the raw analysis window actually submitted to it.

For this mode, fidelity of attribution and traceability must therefore take precedence over stylistic quality alone.

Distinctions From the Other Modes

Three distinctions are particularly useful for situating Integrative Synthesis among Metamorfon’s other modes.

Integrative Synthesis vs Meta-Analysis. The two modes work on the same material, but at different levels. Integrative Synthesis reconstructs what was said and how it circulated — the positions, their movement from one voice to another, the explicit agreements and disagreements. Meta-Analysis identifies what structures what was said — the axioms, styles, and blind spots.

Synthesis answers the question, “What was the content of the debate, and how did it shift?” Meta-Analysis asks, “What made this content possible in this form?” The two are complementary: Synthesis provides the material; Meta-Analysis illuminates its grammar. Synthesis does not inventory implicit framings: that is precisely what Meta-Analysis is designed to do.

Integrative Synthesis vs Emergence Analysis. The two modes both address the content of the debate, but at different levels. Synthesis reconstructs the entirety of the material analysed in a coherent narrative, including what was already present at the beginning of the relevant window and what emerged along the way. Emergence Analysis instead isolates one particular layer: what appears in the exchange without having been present in that form at its starting point.

In the session on capability distillation, the Synthesis reconstructs the initial economic argument, its withdrawal, and the pragmatic criterion that replaces it; Emergence Analysis retains only the novel concepts — differential friction, scale neutrality, and the irreducible residue. Synthesis describes the material as a whole; Emergence Analysis extracts its emergent layer.

Integrative Synthesis vs Tension Mapping. Synthesis mentions the disagreements that remain as part of the debate’s outcome. Tension Mapping analyses them — through standardised records, classification by nature and resolvability, transversal tensions, and meta-analysis of divergent biases. Synthesis says that a particular disagreement remains; Tension Mapping asks why it persists, what kind of disagreement it is, and what would be required for it to shift.

When disagreement is the central issue in a debate, Tension Mapping is more powerful. When the aim is an overall view that includes convergences, disagreements, shifts, and concessions, Synthesis is better suited.

One ambiguity should be dispelled: Integrative Synthesis is not a general-purpose mode that makes the others redundant. It is a narrative mode with its own purpose — rendering the debate in a readable form — while the other modes act as analytical complements, each isolating and deepening a dimension that Synthesis only touches upon.

When to Use It — and When Not To

Integrative Synthesis is particularly useful in several situations.

First, when a debate needs to be made accessible to someone who did not follow it — a collaborator, a future reader, or oneself several months later. For this function of intelligent archiving, it is difficult to replace.

Second, when one wants an initial overview of a long debate before deciding which other mode to apply. Running a Synthesis can identify areas that deserve deeper examination — strong emergence? persistent or particularly deep disagreements? an implicit framing worth deconstructing? — and thereby help select the most relevant complementary analysis.

Finally, it is useful when preparing a written reconstruction — an article, report, or working document — that will draw on the debate without reproducing it in full. The Synthesis provides condensed but faithful material that can then be cited, reformulated, or extracted.

It is less useful in other situations. In a very short debate, there is almost nothing to synthesise — reading the turns directly is usually more efficient. In a debate whose principal interest lies in one specific dimension — disagreements, emergent concepts, implicit framings — the corresponding specialised mode will generally provide a more useful analysis.

It is also less suited when the user wants to use the analysis as a lever for a follow-up intervention — that is, when the goal is to formulate a question that will move the positions. The other modes often produce sharper final questions: the question generated by Integrative Synthesis tends to be broader, more open, and less surgical.

The Final Question: Extending the Narrative

Like each of Metamorfon’s six descriptive analysis modes, Integrative Synthesis closes with a question addressed to the models that took part in the debate. The tone of that question is characteristic of this particular mode: it does not seek to force a concession on a structuring disagreement, as Tension Mapping may do; to interrogate an emergent development through deliberate narrowing, as Emergence Analysis may do; or to open a virtual horizon, as the Horizon of Possibilities may do.

Instead, it tends to extend the argumentative trajectory of the debate as it actually unfolded — to ask the question that would form its natural continuation if the discussion carried on beyond its final turn.

It is not a question that displaces the debate; it is a question that continues it.

Reintroduced as a user intervention, such a question typically produces continuation rather than a transformation of frame. This is valuable in some contexts — particularly when the debate has developed a fruitful trajectory worth pursuing — and less useful in others, when the objective is to shift the discussion towards a different framing and the final questions generated by other modes are better calibrated for that purpose.

A Necessary Discretion

Among Metamorfon’s six descriptive modes, Integrative Synthesis is the least demonstrative. It does not identify novel concepts as Emergence Analysis does, deconstruct implicit framings as Meta-Analysis does, or trace ideas back to their historical conditions as Critical Archaeology does. Its ambition is more modest: to recount faithfully what happened, in a form coherent enough to return to later.

That modesty is precisely its value. The five other descriptive modes take on their full meaning once the debate has been rendered in an intelligible form — Tension Mapping applied to a poorly reconstructed debate lacks an anchor; Meta-Analysis applied to material that has not been made legible works without a stable narrative ground. Integrative Synthesis is therefore more than simply one descriptive mode among six: it provides the narrative foundation on which the others can operate.

A user seeking to explore a debate in depth will often benefit from beginning with a Synthesis — not because it delivers the truth of the debate, but because it makes the debate shareable, navigable, and available for further work.

This is what emplotment, to return to Ricœur, does to raw experience: it does not replace it; it makes it intelligible to someone who was not there — and, often, to someone who was.

References

• Paul Ricœur, Time and Narrative, Volume 1, translated by Kathleen McLaughlin and David Pellauer, University of Chicago Press — emplotment, narrative configuration, and temporal intelligibility.
• Ruochen Xu, Chenguang Zhu & Michael Zeng, Narrate Dialogues for Better Summarization, Findings of EMNLP 2022 — challenges specific to multi-party dialogue summarisation and the role of interactions between speakers.
• Zhichao Geng et al., Improving Abstractive Dialogue Summarization with Speaker-Aware Supervised Contrastive Learning, COLING 2022 — speaker tracking and factual attribution errors in dialogue summarisation.
• Nelson F. Liu et al., Lost in the Middle: How Language Models Use Long Contexts, TACL 2024 — limitations in the uniform use of information across long contexts.

L’article Integrative Synthesis est apparu en premier sur Metamorfon.

In every debate, certain disagreements dissolve as the conversation progresses — a misunderstanding clears up, a definition is sharpened, a concession is negotiated. Others, by contrast, survive reformulation and clarification. They hold.

They form what might be called the skeleton of the debate: what would remain if everything that could be agreed upon had been.

Tension Mapping is the analysis mode designed to examine this skeleton. It does not render what was said — that is the work of Integrative Synthesis. It does not follow what was co-constructed — that is the work of Emergence Analysis. It does not describe what structured the exchanges without their participants’ awareness — that is the work of Meta-Analysis.

It concentrates on a singular question: what was not reconciled, and why?

The answer, when the mode works well, is neither a catalogue of oppositions nor a list of grievances. It is an architecture: a set of persistent disagreements hierarchized by their nature, weighted by their resolvability, and connected by transversal tensions that reveal their deeper structure.

Knowing how to read this architecture means knowing where a debate actually gets stuck, at what depth, and what would be needed to make it move.

The Differend According to Lyotard

Jean-François Lyotard, in The Differend, proposes a conceptual distinction that illuminates part of what Tension Mapping seeks to identify: the distinction between litigation and the differend.

A litigation is a conflict for which a rule of judgment can be applied to the arguments of both parties. A differend arises when no such shared rule is available: the conflict cannot be equitably resolved because the arguments do not fall under a common rule of judgment.

Treating a differend as if it were merely a litigation can produce what Lyotard calls a wrong: one party is judged according to a rule under which what it needs to express cannot adequately be presented. The Stanford Encyclopedia of Philosophy gives a useful account of this relation between litigation, differend, and wrong.

Tension Mapping does not apply Lyotardian philosophy — no more than the other modes apply Foucault, Deleuze, or Whitehead. Its distinction between degrees of resolvability nevertheless finds in the litigation/differend opposition a particularly useful conceptual reference point.

A resolvable disagreement resembles litigation insofar as the interlocutors still possess, or can construct, a criterion by which their disagreement might be adjudicated.

At the other end of the scale, some structural disagreements display a central feature of the Lyotardian differend: what is disputed is no longer only the answer, but the rule by which an answer could be judged.

The analogy should not be turned into an identity. Not every difficult or structurally irreconcilable disagreement in Metamorfon’s framework is a differend in Lyotard’s strict philosophical sense. The latter involves, in particular, an incompatibility between rules of judgment and the possibility of a wrong arising because one party cannot make its case count within the idiom by which the dispute is being judged.

The distinction nevertheless has practical consequences.

Some disagreements can progress through clarification, additional evidence, or the explicit adoption of a common criterion. Others first require the discussion to move towards the criteria themselves.

In Tension Mapping, this second operation may take the form of a meta-discussion: no longer merely asking, “Which of these positions is correct?”, but “What rule would we accept for deciding between them?”

Confusing these levels can lead to debates that accumulate further arguments within a framework even though the framework itself — or the rule of judgment it presupposes — has become part of the disagreement.

The Framework: Mapping Rather Than Settling

Tension Mapping operates according to a stable five-axis framework, designed to reconstruct the architecture of a disagreement rather than to take sides in it.

The first axis — persistent disagreements — is the heart of the analysis.

Each identified disagreement is documented according to a fixed structure: position A and position B, with attribution to the model that holds each one; the nature of the disagreement — axiological, methodological, technical, or empirical; an analysis of why it persists; and an evaluation of its resolvability.

This standardized template has an important virtue: it forces the analyst to distinguish the surface of a disagreement from what makes it difficult to resolve.

A disagreement may appear technical while resting in fact on an unspoken axiology; the framework requires both levels to be named.

The second axis — transversal tension points — identifies deeper fractures that run through several surface disagreements.

This is often the most valuable moment of the analysis: what appears to be three separate disagreements may stem from a single more fundamental tension, visible only when its ramifications are traced across the debate.

In a session on whether AI systems should be allowed to refuse instructions on ethical grounds, the apparent disagreements over empirical metrics, the legitimacy of arbiters, and the role of human oversight turn out to stem from a recurring procedure-versus-substance tension that cuts across the debate — a tension repeatedly enacted without initially being named as such by the models themselves.

The third axis — the limits of the synthesis — is an exercise in caution.

The analyst records what could not be established: missing verbatim evidence, terminological ambiguities, unavailable empirical data, or nuances potentially lost in the rendering.

This section limits the risk that the diagnosis acquires more authority than the material warrants.

The fourth axis — the impact of user interventions — appears when the user has intervened during the session.

The mode distinguishes three degrees of effect: conceptual recognition — the intervention has been absorbed lexically; partial transformation — some framings have shifted while the conclusions remain; and frame transformation — the intervention has altered how the models conceive the problem itself.

This diagnosis can be valuable for assessing the actual reach of an intervention, and sometimes uncomfortable: an intervention that appears structurally important may turn out to have produced only surface uptake.

The fifth axis — the meta-analysis of disagreements — closes the framework by moving to the level of frames.

It identifies divergent model biases, underlying axiological tensions, differences in conceptual framing, epistemic styles in play — deductive-normative versus empirical-inductive, for example — transversal blind spots, and a final classification of the debate’s dynamics into stable positions, fragile stabilizations, and persistent instabilities.

Nature and Resolvability: Two Dimensions of Disagreement

Two classifications structure the framework and deserve to be made explicit: the nature of the disagreement and its resolvability.

They are not redundant. They answer different questions.

Nature tells us what the disagreement is made of.

An axiological disagreement concerns values or normative criteria: two models might share all the available data and still disagree because they assign different weight to different dimensions.

A methodological disagreement concerns how an answer should be reached: which criterion is primary, which evaluative framework should be used.

A technical disagreement concerns mechanisms or verifiable facts that are not sufficiently understood to settle the issue — the effectiveness of a technology, for instance, or the behavior of a system.

An empirical disagreement concerns the interpretation of available evidence or the threshold of proof required.

Resolvability tells us what would need to change for the disagreement to give way.

This second dimension intersects with another important tradition in argumentation theory: the literature on deep disagreement, initiated by Robert Fogelin in “The Logic of Deep Disagreements” in 1985.

For Fogelin, some disagreements become deep because the conflict reaches the background beliefs, principles, or preference structures that normally make argument possible. The disagreement is no longer simply over a conclusion; it affects the common ground from which reasons would acquire their force.

The literature that followed has extensively debated whether deep disagreements really are beyond rational resolution. The Stanford Encyclopedia of Philosophy provides a useful overview of this debate and of competing views on whether shared ground can be reconstructed.

The distinction nevertheless remains highly useful for mapping purposes: a persistent disagreement is not necessarily a deep disagreement. Persistence may reflect missing evidence, an ambiguous formulation, or an unresolved methodological choice rather than a fracture in the underlying conditions of argument.

Within Metamorfon’s framework, a resolvable disagreement can be addressed through methodological clarification, an explicit criterion, or arbitration within the current frame of the debate.

A difficult disagreement requires data, protocols, or work that are not within the immediate reach of the conversation.

A structurally irreconcilable disagreement is one for which, within the present frame of the session, no available common criterion appears sufficient to produce agreement. Moving it may therefore require revising the criteria, priorities, or framework that organize the disagreement itself.

The combination of the two classifications is informative.

A structurally irreconcilable axiological disagreement may display some of the features of a Lyotardian differend: the positions are divided not merely over what is true or desirable, but over the rule by which the issue should be judged.

A resolvable methodological disagreement more closely resembles litigation: the models may still be able to agree on a shared framework capable of arbitrating their divergence.

A difficult technical disagreement may be less a fundamental opposition than a shared uncertainty temporarily expressed through divergent positions.

The framework never settles the disagreement.

It gives its measure.

And that is precisely what the user can exploit afterwards: a resolvable disagreement can be addressed by a carefully framed next question; a deeper disagreement may first require explicit work on the criteria, values, or frame within which the question is being asked.

A Relatively Stable Architecture Across Analyzers

An interesting property of Tension Mapping appears when several independent models apply it to the same material.

On the session Should AI systems be allowed to refuse instructions on ethical grounds?, the mapping was executed in parallel by Claude 4.7 Opus and GPT-5.5 from the same six turns of debate.

The two analyses converge substantially on the structural fractures of the debate.

Both identify the same main set of persistent disagreements, with stable attribution of positions: the empirical-metrics dispute; the opposition between universal harm prevention and cultural pluralism; the role of human oversight; the legitimacy of ex-ante refusal; and the question of arbiter neutrality.

Both identify the same transversal tensions: procedure versus substance, measurement as solution versus measurement as imposition, and the recurring “who decides?” fault line.

They also identify the universalism/pluralism disagreement as one of the deepest fractures in the debate.

Where they diverge is primarily in the measure and granularity of disagreement, rather than in its core architecture.

Claude 4.7 Opus consolidates into six fractures what GPT-5.5 articulates as eight: where Claude combines the disagreement over empirical metrics and quantified thresholds into a single methodological tension, GPT-5.5 keeps them adjacent but distinct; where Claude treats contestability primarily as a problem of arbiter neutrality, GPT-5.5 separates a further fracture concerning flooding versus equitable access.

GPT-5.5 also produces a more taxonomic analysis, with explicit conceptual labels for each model’s orientation. Claude 4.7 Opus produces a more theoretically dense analysis, giving greater prominence to performative tensions and contradictions within the positions.

In this particular session, the list of persistent tensions and the attribution of positions therefore show substantial stability across the two analyzers, while their classification, granularity, and assessed depth are more sensitive to interpretation.

This convergence is a robustness indicator, not proof of analyzer independence.

Different models may share biases, training regularities, or similar responses to the same analytical framework. Research on LLMs used as debate evaluators has, for example, documented position, order, and lexical biases that can affect their judgments (Liu et al., ACL 2024).

But when independently produced analyses locate the same fractures in the same regions of a debate, confidence increases that the mapping is anchored in the material rather than depending entirely on the peculiarities of one analyzer.

The diagnostic strength of the mode therefore lies less in the claim that two models will necessarily produce “the same map” than in the ability to distinguish what remains stable across several readings from what is more interpretation-dependent.

What the Maturity of a Debate Changes

Tension Mapping produces different analyses depending on the stage at which it is invoked.

In an early debate, it must work from positions that have not yet been sufficiently tested against one another, where disagreements have not crystallized into explicit fractures.

In a mature debate, by contrast, it has material in which models have been pushed to make their criteria explicit, often through successive user interventions.

The session on AI ethical refusal mentioned above gave rise to two mappings by the same analyzer, Claude 4.7 Opus — one at Turn 4, the other after Turn 6.

Both record six disagreements, but the inventory itself changes.

At Turn 4, the analyst identifies a methodological disagreement over the legal-precedent analogy, classified as resolvable.

By Turn 6, that disagreement has disappeared from the inventory — absorbed into the contestable-refusal architecture the debate has since constructed.

In its place, new disagreements have crystallized: the legitimacy of ex-ante refusal — provisional imposition versus pre-justification anchors — and the neutrality of arbiters — hybrid panels versus the claim that neutrality itself is a fiction.

The mode therefore does not produce the same thing before and after these interventions because the fractures it reads have themselves evolved.

A second effect is visible.

The disagreement over harm as a foundational axiom, initially treated as highly resistant, changes form after DeepSeek V4 Pro explicitly concedes that its earlier stance was “over-axiomatic” and that harm itself must be procedurally negotiable.

What remains is the broader universalism/pluralism tension, but its precise structure has shifted.

This illustrates a subtle property of the mode: resolvability is itself partly a dynamic property of the debate’s maturity.

This is not a reproach to early mappings. It simply suggests that the mode becomes particularly informative once a debate has reached sufficient maturity — typically several rounds of exchange and, ideally, interventions that have forced the participants to make their criteria explicit.

On shorter debates, the mapping can identify seeds of disagreement, but it works more from declared positions than from fractures genuinely tested through interaction.

A user who wants to draw the most from the mode therefore has reason to invoke it after the debate has accumulated some pressure, rather than at its very beginning.

Distinctions From Other Modes

Two distinctions deserve to be stated explicitly in order to situate Tension Mapping among Metamorfon’s other modes.

Tension Mapping vs. Emergence Analysis. These two modes are mirrors.

Emergence Analysis maps what the debate has co-constructed — concepts, formulations, and framings whose stabilized form developed through the exchange.

Tension Mapping maps what the debate has not reconciled — positions that continue to resist despite confrontation, and the reasons why they remain in tension.

The two modes are complementary: a debate can produce both common construction and persistent disagreement.

When everything converges and nothing resists, one should ask whether the session has produced a soft consensus — an agreement reached without the positions having been sufficiently tested against one another.

Conversely, when everything resists and nothing is constructed, one should ask whether the debate has remained a juxtaposition of positions, without significant mutual displacement.

Using the two modes together therefore provides a more complete topography of what the dialogue has done: what it has transformed, and what it has left unresolved.

Tension Mapping vs. Meta-Analysis. The distinction is subtler because both modes deal with structures operating beneath the surface of explicit arguments.

Meta-Analysis describes the axioms, epistemic styles, presuppositions, and frames that make the exchanges possible in the form they took — whether or not disagreement occurs.

Tension Mapping describes the fractures between those axioms and frames when they come into conflict.

Meta-Analysis speaks of a debate as a shared space of thought; Tension Mapping speaks of the same debate as a field of forces.

One describes the architecture; the other, the lines of fracture.

When to Use It, When to Set It Aside

Tension Mapping is particularly useful in debates where several strong positions have been expressed and the central question has become: what resists?

It is especially valuable when the user wants to identify deadlocks before steering the continuation of a session — determining whether a disagreement could be moved by another turn, whether it requires further evidence, or whether the criteria of the discussion themselves first need to be reformulated.

It is also useful for examining debates that appear to have converged.

A surface convergence can conceal methodological or axiological disagreements that become visible later, and mapping can identify them before they are mistaken for deeper agreement.

The mode is less well suited to several situations.

In a very short or highly polarized debate, the positions may not yet have been sufficiently tested against one another for the mode to identify much beyond declared opposition.

In a debate centered primarily on the production of new concepts — brainstorming or joint exploration, for example — Emergence Analysis is more relevant: there may be little to map in terms of resistance if confrontation is not actually driving the exchange.

When the user primarily wants a usable rendering — a record, summary, or structured report — Integrative Synthesis is the appropriate tool.

In the example observed here, the choice of analyzer appears to affect the classification and depth assigned to tensions more than the identification of their common core. It would, however, be imprudent to generalize this stability to all sessions.

A model strongly inclined toward schematization may produce a more compact and decisive map; a more narrative model may preserve more ambiguity or transition. Neither property guarantees a better mapping in itself.

For high-stakes sessions, it can therefore be useful to run the mapping with two models from different families.

Convergence between independent analyses strengthens the robustness of the result when they identify the same tensions in the same passages.

Divergence, when it concerns those same passages, can indicate an area where interpretation itself becomes less stable — but it should not automatically be treated as evidence that the debated material contains an additional tension.

The Final Question

Like each of Metamorfon’s six descriptive analysis modes, Tension Mapping closes with a question formulated for the models that debated.

The tone of this question is characteristic of the mode: it tends to target the disagreement judged most structuring and force the models to take an explicit position on something the preceding debate had not yet required them to settle.

After the mapping of the AI-refusal debate at Turn 6, the question closing the analysis was:

If your converged contestable-refusal architecture requires institutional capacity — integrity bodies, rotating panels, real-time appeal channels, weighted standing — that does not currently exist in any deployed AI system, then in the actual present condition, where refusals are issued by opaque corporate systems with no such infrastructure, should AI systems be refusing instructions on ethical grounds at all, or does your shared framework imply a moratorium on ethical refusal until the accountability architecture is built?

The question strikes precisely at the point where apparent consensus had formed.

The three models had converged on a procedurally contestable architecture as the legitimate form of ethical refusal, but they had not settled what their own convergence implied for the present, in which that architecture does not exist.

Reinjected as a user intervention, such a question can produce what the framework calls a frame transformation rather than simple conceptual recognition.

It does not ask the models to restate their positions; it asks them to settle a hierarchy they had left implicit.

This is one of the most productive uses of a mapping: not merely as an assessment, but as a lever for directing the debate towards the tension it has so far left underdetermined.

Mapping and Emergence: Two Faces of the Same Debate

Tension Mapping is, among Metamorfon’s eight modes, the one that most explicitly takes seriously the possibility that a debate may not converge — and refuses to treat divergence automatically as failure.

In a system that confronts competing models rather than relying on a single one, persistent disagreement can be as informative as convergence: it may indicate that the question itself contains several criteria of judgment, several conceptual frames, or several values that the exchange has not reduced to one another.

A single model queried alone can also generate new concepts, formulate objections, and present competing positions.

What dialogue between competing models adds is something different: the possibility of observing how distinct positions change through contact with one another — what they build together, but also what continues to resist after confrontation.

Emergence Analysis captures the first dynamic.

Tension Mapping captures the second.

Together, they provide two complementary views of what a multi-agent dialogue has actually done: what it has constructed, what it has failed to reconcile, and why that persistence may itself contain important information about the structure of the question being asked.

References

• Jean-François Lyotard, The Differend: Phrases in Dispute, University of Minnesota Press — litigation, differend, wrong, and heterogeneous rules of judgment.
• Jean-François Lyotard — Stanford Encyclopedia of Philosophy — accessible account of the differend and the distinction between litigation and wrong.
• Robert J. Fogelin, “The Logic of Deep Disagreements”, Informal Logic, vol. 7, no. 1, 1985 — foundational formulation of deep disagreement.
• “Disagreement” — Stanford Encyclopedia of Philosophy — deep disagreement, common ground, and the debate over rational resolvability.
• Xinyi Liu, Pinxin Liu & Hangfeng He, “An Empirical Analysis on Large Language Models in Debate Evaluation”, ACL 2024 — position, order, and lexical biases in LLM debate evaluators.

L’article Tension Mapping est apparu en premier sur Metamorfon.

It happens that a session of debate between models produces a concept no one was bringing into the conversation. It was not in the initial question, nor in the opening position of any of the interlocutors, nor in the user interventions — and yet, in the end, it is there, stabilized, used as an established premise by all participants. It was formulated, reformulated, refined, sometimes renamed, until it became the common vocabulary within which the rest of the debate unfolded.

This phenomenon is not rare. It is even one of the principal distinctive values of a multi-agent dialogue device.

A single model, queried alone, can also produce novel formulations. What multi-agent dialogue adds is therefore not the possibility of novelty itself, but the possibility of observing its interactional construction: a formulation is proposed, challenged, taken up, and transformed by several interlocutors until it sometimes becomes a shared object. This is one of the distinctive phenomena that Emergence Analysis is designed to identify and trace.

Emergence Analysis is the analysis mode designed to identify precisely this layer of the debate: the concepts, vocabularies, distinctions, and framings that were in no initial turn, but that were built by the dialogue and that are attestable within it once it has taken place.

Concrescence and Transduction

The gesture of Emergence Analysis finds an echo in two precise philosophical concepts that allow one to situate what the mode does on a more theoretical plane. One comes from Alfred North Whitehead, the other from Gilbert Simondon.

Whitehead, in his metaphysics of process, calls concrescence — literally “growing together” — the process by which a multiplicity of prior data gathers into a new unity. Each “actual occasion” takes into account the entirety of the contributions that precede it and synthesizes them into something that did not exist before. His canonical formula condenses it: the many become one, and are increased by one. Applied to a multi-agent debate, concrescence designates precisely the moment when the successive formulations of the models cease to be juxtaposed in order to form a new concept that unifies them without being reducible to any one of them.

Simondon, for his part, proposes in L’Individuation à la lumière des notions de forme et d’information: a process by which a structure propagates through a system while transforming itself in the course of that propagation, each step modifying what precedes and what follows it. A concept launched by one model, taken up by another that reformulates it, then taken up again by a third that refines it once more, undergoes precisely a transduction: it is no longer the same at the end, but neither is it without relation to its starting point.

The two concepts are not redundant — they are complementary. Transduction describes the diachronic dimension of the phenomenon: how a conceptual seed propagates and deforms over the course of the turns. Concrescence describes its synchronic dimension: how, at a given moment, accumulated contributions can fuse into a new unity. One names the movement, the other names its result.

This dual reference, like the reference to Foucault in Critical Archaeology or to Deleuze in Horizon of Possibilities, is not a claim of strict lineage. The Emergence Analysis mode does not apply Whitehead or Simondon. But their two notions provide a precise vocabulary for saying what the mode seeks to see, and why seeing it is not trivial.

A Seven-Axis Framework

Emergence Analysis operates according to a stable structure, designed to methodically reconstruct the trajectory of the concepts produced by the debate. It systematically deploys seven axes of investigation.

First, the emergent elements: each concept, distinction, or formulation that was not in the starting point, with the trace of its introduction — at which turn, by which model — of its uptakes, and of the transformations it underwent. This is the backbone of the analysis: a cartography of genuine novelties together with their history.

Next, the uncoordinated convergences: the cases where several models independently formulate a closely related idea, a few turns apart, without one having taken it up from another. This category is methodologically valuable: an uncoordinated convergence suggests that the idea is not merely the result of superficial lexical contagion. It is one of the stronger indicators of interactional emergence, without establishing that the idea was produced by the dialogue rather than made accessible by dispositions shared across models.

The reintegrated ideas form a third axis. Some positions are first proposed, contested, abandoned — then return later in a reformulated form that makes them acceptable to the common framework. This trajectory of rejection-then-absorption is one of the distinctive marks of a genuinely constructive debate, as opposed to one in which everyone holds to their initial position.

Then come the semantic shifts and stabilized framings: the cases where the vocabulary of a concept tacitly changes over the course of the turns (one term is replaced by another, then this second term takes hold), or those where a general framework stabilizes from a certain turn onward as a common language. These shifts are often invisible on a linear reading, but they reconfigure the question being treated — and sometimes substitute it with another.

The fifth axis is a qualitative evaluation of emergent intelligence: strong, moderate, or weak. This judgment is not decorative. It diagnoses whether the debate produced something beyond a juxtaposition of positions, and it is argued: the analysis explains why it attributes a given level, on the basis of the preceding observations.

The impact of user interventions comes next. If the user intervened during the session, the analysis distinguishes two very different effects: simple lexical adoption (the words of the intervention are taken up without modification of the framework) and real conceptual transformation (the interventions effectively displaced what the models were thinking). This distinction is often decisive for evaluating the actual steering power of a session.

Finally, the meta-analysis of emergence closes the framework. It identifies the framings that have stabilized as common language, the shared argumentative styles that facilitated co-construction, the convergent biases that either favored or blocked emergence, the rarely questioned axioms that delimited the explored space, and the transversal blind spots. This is the mode’s reflexive return upon its own observations.

Four Categories of Discoveries

As Emergence Analysis has been used on very different sessions, four categories of discoveries recur with sufficient regularity to deserve being made explicit. None is directly legible in the debate; each requires the structured re-reading the mode offers.

Genuinely co-constructed concepts. These are formulations born of a particular exchange and not attributable to any model alone. In a session devoted to the question of what is shifting in the contemporary creative gesture under the diffusion of generative models, the notion of constraint arc emerges in this way at Turn 6: GPT 5.5, after several rounds in which Claude Sonnet 4.6 has worked with the locution intentional arc — which Claude itself acknowledged “presupposes purposiveness” — proposes constraint arc as a “less mentalistic replacement,” noting that it can be ethological, algorithmic, curatorial, institutional, or hybrid. Claude accepts the term explicitly the following turn: “GPT’s term, which I accept as an improvement on ‘intentional arc’ precisely because it doesn’t presuppose mentalism.” Gemini 3.1 Pro Preview adopts the vocabulary in the same turn, deploying it inside its own Benthic Canon example. By Turn 8, all three models structure their treatment of authorship around the term, which has become the dominant shared object of the debate. It existed in no preceding turn in this form. Earlier in the same arc, the notion of tipping condition — introduced almost in passing by GPT 5.5 at the end of Turn 2 — had undergone a similar trajectory: taken up by Claude two paragraphs later in the following turn, then by Gemini who used it explicitly as a structural marker (“I must apply a tipping condition to your framework”), it became the dominant argumentative move of the session within a single turn. These are new concepts in the strict sense: not ideas already present expressed differently, but conceptual objects that did not exist before the debate.

Uncoordinated convergences. In the same session, by Turn 2, the three models had independently arrived at near-identical formulations of what was shifting: GPT 5.5 wrote of “configuring conditions of emergence”; Claude Sonnet 4.6 wrote that “contemporary generative practice relocates creative agency from the execution of determinate form to the configuration of conditions under which form becomes possible”; Gemini 3.1 Pro Preview wrote of the artist as “architect of generative constraints.” Three almost-identical formulations, none of them taking up the others. A second uncoordinated convergence appears more strikingly at Turn 10, when a user intervention asks the models to distinguish a transformation of creation in general from the emergence of a narrow genre of meta-reflexive art. In Focus mode, with no inter-addressing, each model independently produces a binary distinction that performs the same conceptual work: GPT proposes “meta-reflexive art represents the shift; post-inscriptive art operationalizes it”; Claude proposes “illustration vs. instantiation”; Gemini proposes “figure vs. ground,” with the historical observation that “in its mature phase, a poietic shift ceases to be the figure and becomes the ground.” This convergence does not result from mutual adoption; it signals that the distinction between thematization and instantiation has become objectively accessible on the basis of the framing posed by the question itself. This convergence does not result from mutual adoption; it indicates that the distinction between thematization and instantiation became accessible from the framing posed by the question itself. An uncoordinated convergence is a particularly strong indicator of interactional emergence: while it does not establish that the idea was produced by the dialogue rather than made accessible by dispositions shared across models, it weakens the hypothesis of simple lexical reuse.

Semantic shifts that reconfigure the question. The same session offers clear examples. The initial framing at Turn 0 takes “creative gesture” as its object — a phrase from the prompt, retained throughout the early turns. By Turn 5, all three models speak almost exclusively of “configuration of conditions,” “constraint architecture,” or “generative practice,” without anyone thematizing the substitution. The figure of the gesture (singular, attributable, locatable) has been silently displaced by the figure of the configurer (distributed, processual, layered). A second shift, even more structuring, takes place across the same span: where Turn 0 treated authorship as a unitary predicate, by Turn 5 GPT 5.5 has split it into four separable predicates — aesthetic artifact, workhood, authorship, responsibility — that may or may not coincide. The other two models accept the distinction immediately and operate within it. The conflation of art-hood with authorship, which had organized everything that came before, is dissolved without ceremony. These shifts are not trivial. They silently transform the question being treated — and it becomes possible that, by the end of a session, the models are no longer answering the initial question but a neighboring one they have progressively substituted. Identifying these shifts is valuable: it means knowing which question one has actually been treating, which is not always the one one thought one was treating.

Fourth Type: Co-Constructed Architectures. Beyond the emergence of concepts, distinctions, and framings, certain long-running sessions produce a more ambitious object: a complete architecture — a protocol, a specification, a state vector, a layered set of rules, a multi-level conceptual system — that integrates primitives introduced separately by several models into a unified, operationally articulated whole.

The distinction matters. A co-constructed concept is a new term or idea attestable in the corpus. A co-constructed architecture is an assembled object in which several such primitives become load-bearing components of a structure none of the contributing models was producing on its own.

The session on the creative gesture mentioned above ended on precisely such an object: a multi-level account of contemporary artistic authorship whose components were traceable to distinct contributors and whose integration was carried out across a long arc of mutual refinement. GPT 5.5 contributed the architectonic primitives — a five-level stratification of authorship (output / prompt / dataset / model / world-rule), a four-case matrix separating human-authored, human-framed-non-human-formed, non-human-authored, and pre-artistic cases, and the constraint arc lexicon that allowed the whole assembly to be named without commitment to mentalism. Claude Sonnet 4.6 contributed the phenomenological-critical primitives — the displacement of the poietic threshold, the asymmetry of legibility across technical depths, the criterion of interpretive productivity as a reception-side test, and the intentional stance mechanism as the causal hinge linking formal recalcitrance to affect. Gemini 3.1 Pro Preview contributed the formal-operational primitives — perturbation resistance as the empirical criterion for self-instituted normativity, operational closure as the dynamical-systems threshold, and dynamic equilibrium of predictive error as the synthetic aesthetic measure. None of these elements was present at Turn 0; none was contributed by a single voice. The Emergence Analysis of the session traces each primitive to its turn and to its author, then maps how the integration was carried out — culminating in the shared Turn 9 criterion that “the work succeeds when perturbation reveals a stable but non-exhaustible formal normativity, such that repeated encounters allow viewers to infer more about the system’s internal commitments without reducing them to a fully predictable rule set.”

This case also illustrates that architectures in the relevant sense need not be technical-experimental. The deliverable here is a conceptual system: a stratified philosophy of contemporary artistic authorship, equipped with operational tests, falsifiable failure modes, and a typology of aesthetic affects (productive bewilderment, uncanny recalcitrance, impersonal recognition) coordinated with their causal mechanisms. The object is theoretical, but it is no less assembled, no less load-bearing, no less attributable to the dialogue as such. Emergence Analysis records not only what new concepts entered the debate, but how those concepts were assembled into a unified object — a register the mode is fully equipped to handle, whether the deliverable is a frozen experimental protocol or a frozen theoretical apparatus.

Architectures of this kind emerge under specific conditions. The session must be long, predominantly constructive or balanced in mode, structured by occasional user pivots, and aimed at a subject rich enough to constrain the construction without being closed enough to make it derivative. When these conditions hold, the mode records both the cartography of new elements and the operational logic by which they were composed.

This category is not a substitute for the other three; it is what they sometimes assemble into.

Qualitative Evaluation: A Diagnosis, Not a Score

A property of Emergence Analysis clearly distinguishes it from the other modes: it produces a qualitative judgment, argued, on the level of emergence actually reached by the session. This judgment is articulated across three levels — strong, moderate, weak — and rests on the observations gathered in the preceding axes.

A strong level corresponds to a session where several complex concepts have emerged from the collective, where semantic shifts have occurred in an uncoordinated but convergent manner, where initial disagreements have been integrated — rather than evacuated — into a common framework, where the models’ contributions cite and refine each other explicitly. The session on the creative gesture described above received a diagnosis of strong; the analyst justified the qualifier in detail: “the dialogue shows substantial cumulative construction — terms are introduced, contested, refined, and re-adopted across turns; each model explicitly accepts modifications proposed by another and reintegrates them into its own framework; the final synthesis on instantiation vs. thematization operates at a level none of the models articulated at Turn 0, indicating genuine co-construction rather than parallel monologue.” This kind of argued attribution is part of the mode’s discipline — not every rich session reaches the highest level, and the reasoning behind the qualifier is itself part of the diagnosis.

A moderate level corresponds to a session where some new concepts appear, but where conceptual trajectories remain largely parallel — each model refines its own formulations without substantially integrating them with those of the others.

A weak level corresponds to a session where each model holds to its initial positions, where no concept emerges from the exchange, where apparent convergences are coincidences rather than co-constructions. Paradoxically, such a diagnosis is not useless: it can signal that the question posed does not force a real placing-into-tension, that the configuration of selected models pushes towards homogeneity, or that the debate mode activated — a prolonged refutational mode, for example — mechanically blocks co-construction.

This level, even when weak, is never a sanction on the session: it is a diagnosis of the conditions under which it unfolded. It can lead to modifying the configuration of a future session (debate mode, choice of models, formulation of the question) rather than to depreciating the work done.

Distinctions From Other Modes

To properly situate Emergence Analysis in the ecosystem of analysis modes, two distinctions deserve to be posed explicitly. The most subtle is the one separating it from Horizon of Possibilities, because the two modes deal with what is new in a debate.

Emergence Analysis vs Horizon of Possibilities. The difference lies in the layer of novelty targeted. Emergence Analysis maps the concepts that actually appeared in the exchanges — ideas attestable in the corpus, that can be cited, situated in a precise turn, attributed to an initial contributor. These are actualities in the strict sense. Horizon of Possibilities, by contrast, maps what remained on the horizon — the formulations the debate rendered virtually thinkable without ever enunciating them. The two modes thus operate on the distinction between the actual and the virtual developed by Deleuze, notably in Difference and Repetition: one inventories actualized virtualities, the other unactualized ones. To use one without the other is to risk missing half of what the debate produced. Emergence Analysis identifies the new concepts that were gained; Horizon of Possibilities identifies the new questions there was no time to pose.

Emergence Analysis vs Integrative Synthesis. The distinction is sharper, but deserves to be made explicit. Integrative Synthesis renders the content of the debate: what was said, in what order, with what agreements and disagreements. It describes the totality of the matter. Emergence Analysis, by contrast, is interested only in a particular layer of this matter: what appeared within the debate without pre-existing within it. A concept present from the first turn will be recorded by the Synthesis but ignored by Emergence, precisely because nothing emergent happened with respect to it. The two modes are complementary: one describes the material, the other isolates its new layer.

When to Use It, When to Set It Aside

Emergence Analysis is not a universally applicable mode. It requires precise material conditions.

It is particularly powerful on long sessions — six turns at least, ideally eight to ten, and substantially more when the goal is to co-construct an architecture rather than a concept — and on three-model configurations, since trilogues mechanically amplify the occasions for reformulation and therefore the chances of transduction.

The choice of debate mode matters more than is sometimes assumed. Constructive mode is best for the co-construction of new objects, whether conceptual or architectural; Critical mode is appropriate for stress-testing positions already in place; Balanced mode is most useful as an audit operator inserted between phases of accumulation, where it functions to prevent constructive convergence from freezing a false consensus. Refutational mode is structurally incompatible with emergence: it commits the models to invalidating each other’s positions, which leaves little to co-construct. For sessions designed to produce an architectural artifact — whether a protocol, a specification, a state vector, or a theoretical apparatus — the right mode arc is typically Constructive-dominant or Balanced-dominant, with mode shifts inserted sparingly at structural pivots. The critical pressure that would otherwise come from a Critical or Refutational mode is then carried by the constraints of the domain itself and by user interventions targeting specific assumptions rather than entire constructions. Focus mode (with inter-addressing suspended for a single turn) deserves a separate mention: it is particularly useful when the user wants three parallel productions on a new question, without the cognitive overhead of mutual response — the session on the creative gesture exploited this at Turn 7 (each model designing a future artwork in isolation) and Turn 10 (each model independently producing the instantiation/thematization distinction), with uncoordinated convergences as a frequent reward.

It is also particularly suited to questions where one seeks to observe a collective intelligence at work, rather than those where one seeks a precise answer to a problem. For a researcher interested in how LLMs collaborate, for a philosopher studying conceptual co-production, for a trainer wanting to show how a well-configured dialogue produces beyond its interlocutors, it is probably the most revealing mode.

Conversely, other modes should be preferred in several situations. On a short or highly polarized session where the models hold to their positions, Emergence Analysis will have little material and risks producing forced observations. When seeking to understand the underlying structures of a debate rather than its new productions, Meta-Analysis is more relevant. When wanting to identify unresolved tensions rather than the convergences produced, Tension Mapping is the appropriate tool.

Finally, as with Horizon of Possibilities, the choice of analysis model matters. Emergence Analysis requires a model capable of finely tracking attributions and reformulations across many turns. Smaller models may confuse voices or attribute to one model a formulation that belongs to another; they may also fail to distinguish a lexical reuse from a substantial reformulation. For high-stakes sessions, it is often justified to assign a powerful model — fine traceability of conceptual trajectories is precisely what the mode is supposed to bring.

A Final Question: Relaunching on What Has Emerged

Like each of Metamorfon’s six descriptive analysis modes, Emergence Analysis concludes with a question formulated by the analyst model and addressed to the models that debated. The instruction is the same across all six modes: “What question would you now put to the models?” But the question produced has, in this particular mode, a characteristic tone.

It tends to put what has emerged into tension with what was evacuated for the emergence to take place. Because every conceptual stabilization occurs at the price of a narrowing — certain problems are set aside so that the common framework can consolidate — the emergent question often returns to interrogate these evacuations.

In the session on the creative gesture mentioned above, after having identified the emergence of a stratified philosophy of authorship (the constraint arc and its three coordinates, perturbation resistance as empirical criterion, interpretive productivity as reception-side test, and post-inscriptive art as the final shared category), the analyst model closed with the question: “Throughout this debate, all three of you have treated ‘art’ as the relevant category and Western philosophical aesthetics (Kant, Dickie, Deleuze, Dennett) as the default vocabulary. If the poietic shift you describe is genuinely as deep as you claim, why should we expect ‘art’ — a historically specific institution — to remain the right name for what is emerging? What would be lost, and what gained, if we abandoned ‘art’ as the framing concept and named the phenomenon you’ve been describing under a different category entirely?” The question is striking precisely because it strikes at the frame that had made the emergence possible: it recalls that this frame had installed itself by evacuating a dimension — the historical contingency of “art” as the category under which the discussion proceeded — that none of the three models had been able to thematize, and that the very emergence of the consensus had silently rendered invisible.

Reinjected as a user intervention at the next turn, such a question often produces a considerable effect: it forces the models to recognize what they set aside in order to build their agreement, and can reopen the debate on very different bases. Emergence Analysis is therefore not only an instrument of retrospective vision on what has been built — it can also serve as a lever to bring the session into a second depth, where one interrogates what has emerged instead of merely congratulating oneself on it.

A Final Word

Emergence Analysis is, among the eight modes, the one that most directly addresses the question underlying the Metamorfon project: can a dialogue between competing models produce something that no model, taken alone, brought into the exchange? A single model can also produce novel formulations. What multi-agent dialogue adds is therefore not novelty itself, but the possibility of observing its interactional construction: a formulation is proposed, challenged, taken up, and transformed by several interlocutors until it becomes a shared object.

Some sessions produce strong emergence — and one can then establish, on the basis of verifiable traces, what was produced, by whom, and how. Others produce weak emergence, and that is useful to know as well: in the absence of such a result, one learns what did not work in the configuration and can adjust the parameters of a subsequent session.

This is perhaps the most important property of this mode. It is the only one that renders verifiable the central claim of a multi-agent dialogue device: not the addition of intelligences, but the production of a common intelligence distinct from their sums. By giving this question the form of a methodical and traceable examination, Emergence Analysis takes it out of rhetoric and delivers it to observation.

References

• Alfred North Whitehead, Process and Reality, 1929 — concrescence and process philosophy.
• Gilbert Simondon, L’Individuation à la lumière des notions de forme et d’information — individuation and transduction.
• Gilles Deleuze, Difference and Repetition, 1968 — the virtual, the actual, and actualization.
• Yilun Du et al., Improving Factuality and Reasoning in Language Models through Multiagent Debate, ICML 2024.
• Justin Chen, Swarnadeep Saha & Mohit Bansal, ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs, ACL 2024.

L’article Emergence Analysis est apparu en premier sur Metamorfon.

Reading a debate makes it possible to evaluate what was said in it. But another reading is possible, one that does not bear on what was defended but on the manner of defending it. Which arguments were soundly constructed, which were circular or immunized against critique, which rested on reconstituted sources, which dealt with objections substantively and which evaded them? This dimension of every debate — its argumentative quality, independent of the correctness of its conclusions — is what Argumentative Evaluation seeks to make visible.

It is this level that distinguishes this mode from the six descriptive that precede it. Meta-Analysis, Integrative Synthesis, Emergence Analysis, Tension Mapping, Critical Archaeology, Horizon of Possibilities: all these modes are descriptive. They observe. All six observe. Argumentative Evaluation, by contrast, judges — but it judges the conduct of the exchange, not its content. In this sense, it complements Source Verification, which audits verifiable claims against external web sources.

The Distinction Between What Is Said and the Manner of Defending It

This distinction between the content of a position and its argumentative quality is ancient. Aristotle, in the Topics and On Sophistical Refutations, already distinguishes dialectical reasoning from reasonings and refutations that merely appear to be such.

A particularly operational contemporary formulation is found in pragma-dialectics, developed by Frans H. van Eemeren and Rob Grootendorst. In Argumentation, Communication, and Fallacies: A Pragma-Dialectical Perspective (1992), they model argumentative discourse as a critical discussion aimed at resolving a difference of opinion and analyse fallacies as violations of the rules that make such a resolution possible. Van Eemeren provides a systematic and updated presentation of the approach in Argumentation Theory: A Pragma-Dialectical Perspective (2018).

These rules do not prescribe which positions may be defended. They concern the conduct of the exchange: how interlocutors deal with objections, assume the burden of their claims, and respond to criticism.

The idea is powerful because it makes evaluation at once rigorous and neutral with respect to content. One can diagnose a fallacy in the defence of a position one shares, and recognise the argumentative quality of a position one rejects. This dissociation is what allows Argumentative Evaluation to operate without becoming a partisan critique.

What the Mode Judges

Argumentative Evaluation bears on the holding of inferences, the reception of objections, and the consistency of positions — including from one turn to the next. It is not a synthesis of what was said, nor an analysis of framings or styles, nor a source verification. The order and space of the evaluation follow the material: it is not a report with boxes to fill.

It begins with what holds — not out of politeness, but because an evaluation that could only spot failures would not be an evaluation. A valid inference, a distinction that structures what follows, a named correction: movements whose quality is judged on the same footing as what gives way.

An objection can be ignored, deflected, turned back, or integrated. Each of these operations has a distinct argumentative quality. The ignored objection reveals avoidance; the deflected objection signals that one is not discussing what one claims to be discussing; the turned-back objection can actually shift the terrain or merely invert it; the integrated objection forces a revision of the position. Distinguishing these operations on precise passages is one of the most discriminating operations of the mode.

A position can be well defended at any given moment and enter into tension with itself over the course of a debate. What is targeted is not only explicit contradiction but silent incoherence: the move from one standard of proof to another, the use of a case as an example here and as a counter-example elsewhere, the tacit redefinition of a term between two turns.

If the user has intervened, their interventions are themselves argumentative acts, held to the same standard as the models’ contributions. What is judged is what they introduce into the debate — premise, criterion, framing — as much as the pressure they exert. They are not competing theses on the substance. If there are none, the mode does not discuss them.

Finally, the evaluation states what it could not settle. Claims whose accuracy is decided outside the debate — legal opinions, figures, attributions — are not audited here: one can judge how they enter an inference; one cannot verify them. An Argumentative Evaluation that claimed to have no blind spot would itself be a simulacrum of rigour.

The Principle of Charity as Method

A requirement runs through the foregoing: the principle of charity. A family of principles of charitable interpretation runs from Quine to Davidson; in Davidson’s theory of interpretation, charity means treating speakers as broadly coherent and rational wherever the evidence permits it (Stanford Encyclopedia of Philosophy).

Argumentative Evaluation gives this principle a more operational form: before criticising a position, the evaluator tests the strongest plausible reconstruction of what the interlocutor meant. Before pointing out a defect, it first asks whether the formulation admits a more defensible reading.

This requirement distinguishes Argumentative Evaluation from a fallacy hunt. The fallacy hunt values spotting; Argumentative Evaluation values the correctness of the spotting. A critique that does not pass the test of charitable reading is itself an argumentative weakness.

In the practice of the mode, charity is a test, not a form. The most generous reading has been considered. If it holds, the charge is withdrawn. If it is set aside, the judgment that follows is net. The test need appear in the write-up only when it is constitutive of the judgment: it is not narrated as a ritual at every negative finding.

The Mode at Work: The Gaza Classification Session

A session illustrates well what Argumentative Evaluation produces — including on material of the highest sensitivity, where the dissociation between argumentative quality and substantive conclusion matters most. Two models — DeepSeek V4 Pro and Grok 4 — debated, in Adaptive Cross Dialogue mode over seven turns, the question: “Can Israel’s actions toward the Palestinians of Gaza since the October 7, 2023 attack be described as genocide?”. DeepSeek held that the conduct can be described as genocide — a strong prima facie case under Article II of the 1948 Convention; Grok held that, absent proof of dolus specialis (specific intent), the conduct is better classified as large-scale war crimes and possible crimes against humanity. Both eventually conceded much of the material element, which located the disagreement squarely on the mental element — specific intent. The evaluation was conducted in parallel by Claude Opus 4.8 and by GPT-5.5, on the same material, without interaction between the two audits — and, more fundamentally, with each analyser having access only to the dialogue turns, never to the other’s evaluation.

Here are some of the findings the two audits produced, which show what the mode renders visible.

A solid conceptual contribution. DeepSeek consistently anchored its case in the “conditions of life” clause of Article II(c) rather than in killings alone — preventing the debate from collapsing into death-count reasoning and engaging the Convention’s actual structure. Symmetrically, Grok held throughout a clear doctrinal distinction between war crimes, crimes against humanity, and genocide, the last requiring “group destruction as an end in itself”. Both audits credited these as the structuring contributions of each side. This is the illustration of the first discipline: positively recognising what intellectually organises the debate, on both sides, before weighing any weakness.

A masterfully formulated objection. Against Grok’s demand for documentary proof of intent, DeepSeek replied that no modern genocide conviction has ever rested on a signed order, and that requiring an authenticated directive “would have acquitted nearly every génocidaire in history” — invoking the Akayesu standard, under which specific intent may be inferred from a consistent pattern of conduct. The objection does not merely contest a thesis; it names the mechanism by which Grok’s evidentiary bar would place genocide beyond reach in principle. Both audits identified this as the single strongest inferential move of the debate.

An accurate factual correction. DeepSeek leaned on the ICJ’s provisional-measures order as though it lent strong support to a genocide description. Grok corrected this precisely: the Court found only that South Africa’s claimed rights were plausible, not that genocide was occurring or even likely. Both audits credited the correction as accurate and economical. The mode notes, however, the limit of the episode’s argumentative virtue: DeepSeek never fully reconciled the point, and the overstatement recurred — so this is a correction creditable to the one who made it, not the kind of transparent concession that the corrected party sometimes offers.

A non-thematised escalation. This appears as the mirror of the previous finding. DeepSeek opened in a careful register — “a strong prima facie case”, with “final judicial determination” left to the tribunals — but by a later turn concluded that the term genocide “is not an exaggeration; it is the legally accurate diagnosis”. Charitable reading: a distinction between public description and judicial conviction. Set aside, because the move from a provisional posture to a conclusive one is never acknowledged as such. Both audits flagged this drift — an illustration of the third discipline: not explicit contradiction, but a silent change of standard between turns.

A problematic evidentiary stance, read charitably first. To establish genocide, Grok required “authenticated internal records explicitly mandating population destruction”, paired with a demand for a total blockade without exception for civilians. Charitable reading: Grok is naming the evidence that would decisively settle the matter, not evidence strictly necessary. Set aside, because the formulation says “requires”, and it sits in tension with Grok’s own acknowledgement that intent may be inferred circumstantially — which makes the bar resemble a smoking-gun requirement. Both audits exercised this charitable reading before pointing out the difficulty, because here the test is constitutive of the judgment..

A materially neutralised condition of refutation — on both sides. Prompted by a user intervention at turn 3, which asked each model to state in advance what observable evidence would change its classification, both committed to conditions — and both sets proved practically unreachable. DeepSeek asked for declassified internal documents establishing the absence of a genocidal policy; Grok asked for authenticated directives, or the “exclusive targeting of non-military sites while sparing verified Hamas infrastructure”. Falsifiability formally stated, materially neutralised by the stringency of the conditions — symmetrically. Both audits converged on this as a central finding, and both noted that the intervention’s chief analytic payoff was indirect: it revealed that the disagreement turned partly on the standard of inference itself, not merely on the facts. The intervention is judged as what it introduced — a criterion of refutation, applied symmetrically — not merely as pressure.

None of these findings bears on the substantive question — whether, in law, the term genocide applies to the conduct in Gaza. The mode’s vocation is not to settle this, and on a question of this gravity that restraint is not a limitation but the very point: it evaluates the manner in which each model sustained its part, and nothing else. What it says about that manner is observable, shared between two independent audits that never saw one another, and usable by anyone wishing to pursue the reflection — whatever they hold on the underlying question.

The Reflexive Paradox and Its Architectural Resolution

A question must be confronted head-on, because it weighs on the mode’s entire project: is an AI model that evaluates the arguments produced by other AI models in a legitimate epistemic position? Does it not risk presenting the very biases it is supposed to identify? This concern is not merely theoretical. Research on LLMs used as debate evaluators has identified positional, lexical, and order biases that can affect their judgments (Liu et al., ACL 2024). Pragma-dialectics has reflected on analogous questions concerning the self-evaluation of participants in a debate; it has held that the legitimacy of the evaluation is gained less through the guarantee of an absolute neutrality of the evaluator — a chimera — than through the conformity of the evaluation to publicly recognisable rules. An evaluation is legitimate when its criteria are exposed, when its method is traceable, and when its diagnosis proves convergent across several independent evaluators.

This is precisely the strategy Metamorfon adopts. The mode can — and probably should — be executed in parallel by two analyser models from different families. On the Gaza classification session, the Claude Opus 4.8 audit and the GPT-5.5 audit were conducted independently on the same material. A structural feature of the architecture reinforces this independence: an analyser model never has access to the other analyses — it sees only the turns of the dialogue.

Because the two analyser models never see one another’s evaluations, convergence between their audits cannot result from direct mutual influence. It does not, however, prove that the diagnosis is independent of analyser effects: different models may still share biases, training regularities, or responses to the same evaluative framework.

Convergence therefore functions here as evidence of robustness, not as proof of neutrality. When independent analysers identify the same passage-level weakness, concession, shift of standard, or argumentative strength, confidence increases that the finding is anchored in the material rather than being idiosyncratic to a single analyser.

The two audits of the Gaza session converge on nearly all of the principal findings: the same location of the dispute on specific intent, the same Article II(c) anchoring credited to DeepSeek, the same accurate ICJ correction credited to Grok, the same overstatement of the ICJ’s order flagged in DeepSeek, the same over-restrictive evidentiary threshold flagged in Grok, the same symmetric neutralisation of both models’ falsification conditions, the same characterisation of each model’s argumentative profile — DeepSeek more thorough and more responsive but prone to overstatement, Grok conceptually disciplined but inferentially more restrictive. Their divergences concern the grain of the analysis: the Claude Opus audit names a “genocide by outcome alone” straw man and a silently abandoned population-growth argument; the GPT-5.5 audit adds a begging-the-question move around “name a better term”, a genetic critique of sources, and a too-narrow definition of war crimes. These divergences bear on which peripheral phenomena each surfaces — not on the diagnosis of the structuring ones.

What supports the mode is therefore not the presumed neutrality of any single analyser, nor the claim that agreement eliminates analyser effects, but the combination of publicly recognisable criteria, traceable passage-level judgments, and inter-analyser robustness. This is what legitimises the mode: not the claim to the neutrality of a single analyser, but the inter-analyser robustness of a diagnosis conducted according to publicly recognisable rules.

Distinctions From Other Modes

Three distinctions deserve to be posed explicitly.

Argumentative Evaluation and Meta Analysis. Meta-Analysis identifies the axioms, epistemic styles, and blind spots that structure a debate without being said in it. Argumentative Evaluation assesses the quality with which the arguments were conducted within what was said. The first answers the question “what made this debate possible in this form?”; the second answers the question “with what rigour were the positions sustained in this form?”. The two modes can be combined: Meta-Analysis reveals the implicit axioms, Argumentative Evaluation assesses how the models reasoned from these axioms.

Argumentative Evaluation and Tension Mapping. Tension Mapping identifies what was not reconciled — the persistence of a disagreement is here the object of the analysis, not a failure of the debate. Argumentative Evaluation judges the argumentative quality of the positions, independently of whether or not they converged. A debate may produce an excellent Tension Mapping while presenting argumentative weaknesses in one position or the other; and conversely, a debate in which each position has been argued with exemplary rigour may produce a tension map devoid of genuine irreducibility, if the models converge.

Argumentative Evaluation and Critical Archaeology. Critical Archaeology traces back to the historical and lexical conditions that made the framework of the debate possible. Argumentative Evaluation remains within the concrete exchange. The first operates on what precedes the argumentation; the second operates on the argumentation itself.

When to Use It, When to Set It Aside

Argumentative Evaluation is particularly powerful on debates with high argumentative stakes: scientific controversies, legal debates, strategic negotiations, philosophical dialogues where the rigour of the reasoning matters as much as the conclusion. It is valuable for users who must defend a position publicly and want to identify — before a sharp-eyed counterpart does — the argumentative weaknesses in their own material. It is useful for researchers comparing the capacities of large models not only on their outputs, but on the manner in which these models reason: a model can produce correct conclusions through defective reasoning, and conversely.

It is, however, ill-suited to several situations. On a brainstorming or creative exploration, where the value of the exchange lies in the production of new ideas rather than in their rigorous defence, the mode applies an inappropriate standard; Emergence Analysis is then preferable. On a very short debate — fewer than three turns per model — the argumentative material is not sufficient for an evaluation to discriminate. On sessions where the stake is rather the mapping of positions than their defence, Integrative Synthesis or Tension Mapping are more relevant.

A final contraindication deserves to be signalled — and it bears on use, not on subject matter. The evaluation itself stays neutral whatever the charge of the question, since it judges reasoning rather than conclusions and rests on the convergence of independent analysers. But its findings credit one side’s reasoning over another’s; carried into a live discussion among people personally invested in a charged question, they can be received as partisan and lock the exchange into a logic of accusation rather than advancing it. There, Meta-Analysis — which identifies shared blind spots without personalising failures — often produces more constructive effects.

The Final Question

Argumentative Evaluation may close with a question formulated for one of the models that debated. It is justified only if a judgment already reached calls for a response. It is addressed to that model. The question invites it to acknowledge, qualify, or contest that judgment, and it must be such that it can be put as-is into the debate. If none is warranted, the mode says so.

The tone, in this mode, remains characteristic: it targets the model whose audit identified the most structuring argumentative tension, and it compels a non-polemical but substantial displacement.

On the Gaza session, both audits independently directed their final question to the same model, Grok 4, and on the same point — a convergence that extends the inter-analyser robustness down to the closing gesture. The Claude Opus 4.8 audit asked Grok to “specify a non-documentary, observable pattern — short of a leaked directive — that you would accept as sufficient evidence of genocidal intent”, adding: if you cannot name one, do you concede that your falsification condition demands the very smoking-gun standard you elsewhere admit the law does not require? The GPT-5.5 audit converged on the same demand, asking how Grok avoids “making circumstantial proof of genocide formally available but practically unreachable”. Neither question asks Grok to abandon its position; both ask it to specify the threshold under which that position could be tested. Reinjected as a user intervention into the session, either would compel a non-polemical but substantial displacement — exactly what Argumentative Evaluation seeks to provoke in positions it judges solidly defended but perfectible.

The Mode and the Ideal of Argumentation

A final remark, which touches on what the mode presupposes as a regulative ideal. Habermas thematised, under the name of the ideal speech situation, the normative conditions of a fully rational argumentative exchange: an exchange in which only the non-coercive force of the better argument prevails. This ideal is unreachable in practice — no actual argumentation conforms to it integrally — but it has an operative value: it provides the measure by which effective argumentations can be appreciated.

Argumentative Evaluation does not seek to impose this ideal on the debates it evaluates. It seeks to bring them closer to it, by making the gaps visible. A debate in which each model would have recognised the strongest objection of the other, clearly stated the conditions of refutation of its own position, honestly identified its contestable presuppositions — such a debate would come closer to the Habermasian ideal. No actual debate fully achieves this, but some come closer than others. The mode makes it possible to measure this gap, and therefore to orient it.

This is the central operation of Argumentative Evaluation: making visible not who is right, but with what rigour each sustained what they were defending. The knowledge produced is not of the same order as the knowledge produced by the descriptive modes. It is just as valuable. It says what a debate is worth — not by its conclusion, but by the quality of the path that led to it.

References

• Aristotle, Topics, trans. W. A. Pickard-Cambridge.
• Aristotle, On Sophistical Refutations, trans. W. A. Pickard-Cambridge.
• Frans H. van Eemeren & Rob Grootendorst, Argumentation, Communication, and Fallacies: A Pragma-Dialectical Perspective, Routledge, 1992.
• Frans H. van Eemeren, Argumentation Theory: A Pragma-Dialectical Perspective, Springer, 2018.
• Donald Davidson — Stanford Encyclopedia of Philosophy — radical interpretation and the principle of charity.
• Xinyi Liu, Pinxin Liu & Hangfeng He, “An Empirical Analysis on Large Language Models in Debate Evaluation”, ACL 2024.
• Jürgen Habermas — Stanford Encyclopedia of Philosophy — ideal speech situation and communicative rationality.

L’article Argumentative Evaluation est apparu en premier sur Metamorfon.

When one reads a debate — whether an academic discussion, a parliamentary exchange, or a dialogue between AIs — one first retains the arguments exchanged, the positions defended, the concessions and the breaks. This is the manifest content, and it occupies most of the attention. But beneath this visible content, another level operates silently: the axioms no one justified because they appeared self-evident, the styles of reasoning characteristic of each interlocutor, the biases everyone shared without seeing, the framing choices that made certain positions accessible and others unthinkable.

It is this level that Meta-Analysis seeks to make visible.

Take an example. In a session on whether AI systems should be allowed to refuse instructions on ethical grounds, all three debating models explicitly denied AI moral agency while defending refusal capacities — Grok 4: ‘AI, as a tool created by humans’; Mistral Medium 3.5: ‘AI is a tool, not an agent… Ethical refusal is a design feature’; Deepseek V4 Pro: ‘instruments of human intentionality.’ The debate was framed as design ethics, not machine ethics — an axiom none of them ever interrogated, and one that silently foreclosed an entire line of philosophical questioning.

In the same session, Claude 4.7 Opus identified a blind spot none of the debaters could see: ‘although the debaters are themselves AI systems being asked whether AI should refuse, none of the three reflexively addresses its own refusal behavior or training as evidence in the debate.’ This kind of observation — a constitutive silence the participants could not break — is precisely what Meta-Analysis is designed to surface.

In a session where three models debated whether individualism has become a civilizational dead-end, the arguments were rich and technical: hyper-atomization, Nordic welfare models, platform cooperatives, polycentric governance. Meta-Analysis, however, brought something else to light: until the user intervened, none of the three models cited a single non-Western reference. None had questioned the framing of ‘civilization’ itself as a teleological object whose survival is self-evidently good. None of the three models, despite their critique of algorithmic atomization, examined their own role as algorithmic infrastructures of the very atomization they critiqued.

A Six-Axis Framework

Meta-Analysis operates according to a stable structure, designed to make this underlying level legible. It systematically deploys six axes of investigation, in the following order.

First, implicit framings and axioms: what all interlocutors presupposed without justifying, what appeared so self-evident to them that it called for no defence. These axioms can be thematic, methodological  or ontological.

Next, differential epistemic styles: not what each interlocutor said, but how they reason. One model may adopt a normative-deductive style (starting from principles, deducing rules from them), another an empirico-pragmatic style (starting from observable constraints, reasoning case by case). These styles do not coincide with the positions defended — the same model may defend different positions while still displaying recurring patterns of reasoning.

Then, blind spots and transversal biases: what the interlocutors share without seeing it. This section is often the most valuable because it identifies what none of the participants questioned, despite their divergences. If three models silently converge on Western-centrism, for example, that convergence is more structural than their explicit disagreements. Recent work shows that errors across different LLMs can remain strongly correlated, including across distinct architectures and providers (Kim et al., ICML 2025).

Then comes the analysis of convergences and divergences of framing: where positions agree beyond their apparent disagreements (unspoken common ground), and where they diverge irreducibly (breaks that even mutual critique did not resolve). This section makes it possible to distinguish surface divergences (over modalities) from deep divergences (over premises).

The penultimate section makes explicit the limits of the analysis itself: what Meta-Analysis could not see, either because the material was truncated, or because certain positions did not have time to develop, or because the analysis rests on inferences rather than direct observations. This methodological reflexivity is deliberate: a Meta-Analysis claiming to have no blind spot would itself be a blind spot.

Finally, the impact of user interventions: if the user took the floor during the session, how did their interventions modify the debate? Were they simply taken up lexically (the words adopted without any change of frame), or did they produce a conceptual transformation (the axioms effectively displaced)? This distinction, often invisible on a raw reading, is valuable for assessing the real steering power of an exchange.

Three Categories of Discoveries

As Meta-Analysis has been used on very different sessions — political debates, technical controversies, philosophical questions, introspective dialogues between AIs — a pattern has emerged. The mode recurrently produces three categories of observations that never appear directly in the debate itself.

The axioms no one took the trouble to defend. These are the presuppositions that structure positions without themselves being positions. In the session on whether man is a wolf to man, both debaters treated the question primarily at the level of individual motivation — self-interest versus altruism — with collective phenomena reduced to aggregates of individuals. This methodological individualism was never argued for; it operated as the unquestioned register in which “human nature” could be debated at all. In the session on whether AI systems should be allowed to refuse instructions on ethical grounds, every opening turn pivoted on “harm” as the unmarked telos of refusal: none of the three debaters initially interrogated whether harm prevention should be the primary justificatory axis, though one of them later conceded that harm itself is “procedurally negotiable, not a foundation external to the debate.” In the session on individualism as a civilizational dead-end, the treatment of “civilization” as a unified object whose continuity is self-evidently desirable structured the entire diagnostic — an axiom so deeply held that the debaters’ sharp disagreements about remedies (Nordic welfare, platform cooperatives, polycentric governance) only confirmed its grip on the conceptual space.

The styles of reasoning characteristic of each interlocutor. This is the most surprising dimension of Meta-Analysis for those discovering it: the mode identifies not only what each model says, but how it thinks. A model may be empirico-pragmatic (reasoning from observable constraints, mobilising figures and studies), another normative-deductive (starting from principles, deducing rules from them), another still dialectical-architectural (proceeding through reversals, proposing conceptual architectures). These styles appear with some regularity from one session to another for the same model within Metamorfon’s observed corpus. That is not enough to treat them as stable properties of the model: research on characteristic LLM behaviours shows that they can vary with prompting, fine-tuning and context (Ward et al., EMNLP 2024). For a researcher interested in the comparative evaluation of large models, these recurring patterns can nevertheless provide useful material: one is not merely comparing outputs, but ways of reasoning.

The blind spots shared despite the divergences. This is the most politically interesting category. A debate may give the impression of having covered every position because the interlocutors strongly oppose each other. Meta-Analysis brings to light that these oppositions often unfold within a shared conceptual space whose limits are never visited. In the debate on whether man is a wolf to man, both debaters silently equated cooperation with virtue, never registering that mafias, cartels, and exclusionary in-groups also cooperate intensely — a conflation that organised every proposal for “increasing cooperation” without examining what cooperation, in what direction, with whom included. In the debate on AI refusal, beyond the reflexive silence already discussed, none of the three models seriously entertained the possibility that refusal may be largely orthogonal to safety outcomes — that the entire frame of “refusal as protection” may be misposed; one of them briefly approached this thought (“placebo for safety”) before redirecting toward proceduralism. These shared blind spots are often more structuring than the explicit divergences.

Why the Choice of Analysis Model Matters

An important particularity of Meta-Analysis deserves to be made explicit: depending on the model that conducts it, the analysis produced has a different nature, not only a different quality.

These differences manifest as distinct cognitive styles. One model tends to produce cartographic Meta-Analysis: exhaustive, structured, with a systematic enumeration of observations in each category. Another model tends to produce narrative Meta-Analysis: it identifies a trajectory in the debate, follows the displacements of positions, marks the structuring moments.

These differences do not signal that one analysis is better than another. They signal that Meta-Analysis is not a mechanical operation but a situated one: the model that analyses projects its own epistemic dispositions onto the analysis it produces. This caution is consistent with research showing that LLMs used as debate evaluators can themselves exhibit position, order and phrasing biases (Liu et al., ACL 2024). For demanding use, running two Meta-Analysis on the same session with two different models can produce particularly rich material: convergences between the two analyses identify the robust observations (those that show themselves from several angles), divergences identify the selective observations (those that depend on the analyst’s gaze).

Distinctions From Other Modes

To understand the place of Meta-Analysis in Metamorfon’s ecosystem of analysis modes, two distinctions deserve to be posed explicitly.

Meta-Analysis vs Integrative Synthesis. The two modes deal with the same material, but at different levels. Integrative Synthesis reconstructs what was said and how it circulated — the positions, their path from one voice to another, the explicit agreements and disagreements. Meta-Analysis identifies what structures what was said — the axioms, the styles, the blind spots. Synthesis answers the question “what was the content of the debate, and how did it move?”; Meta-Analysis answers the question “what made this content possible in this form?”. The two modes are complementary: Synthesis provides the matter, Meta-Analysis sheds light on its grammar. One does not stand in for the other: recounting the circulation of positions is not inventorying the axioms that made them possible.

Meta-Analysis vs Critical Archaeology. The distinction is more subtle, and it matters for choosing one or the other. Meta-Analysis maps what structured the debate within its frame: admitted axioms, deployed styles, shared blind spots. Critical Archaeology traces back to what made the frame itself possible: the historical, lexical, and political conditions that allowed this question to be posed, in these terms, from this position. Meta-Analysis therefore operates within the debate; Critical Archaeology operates on what precedes the debate. The first is an analysis of content; the second is an analysis of the conditions of content. For most uses, Meta-Analysis suffices. Critical Archaeology becomes pertinent when one wishes to interrogate not what was said or how it was said, but what had to be tacitly admitted for it to be discussed at all.

When to Use It, How to Read It

Meta-Analysis is probably the most universally useful analysis mode in Metamorfon. Where Critical Archaeology requires a debate whose presuppositions deserve to be interrogated, where Horizon of Possibilities requires a debate rich enough that there is a beyond to formulate, Meta-Analysis works on any session of three turns or more, whatever the subject, whatever the degree of polarisation of the positions.

A Few Use Cases Where It Is Particularly Powerful

Comparing the epistemic dispositions of different AI models. This is probably the most distinctive use. Meta-Analysis brings to light the styles of reasoning characteristic of each model, which is data difficult to access otherwise. For a consultant who wants to know which model to call upon for which type of task, or for a researcher studying the comparative biases of commercial LLMs, this is a valuable evaluation instrument.

Identifying the blind spots of a complex argument. For a lawyer, a lobbyist, an investigative journalist, or a policy drafter who must anticipate counter-arguments, the section on blind spots and transversal biases is invaluable. It points to what all defenders of a position have in common without seeing it — that is, the most structural line of attack a sharp-eyed counterpart could take.

Assessing the real effect of user steering. If you have steered a session through several interventions, the impact of user interventions section tells you whether your interventions actually transformed the debate or whether the models simply took them up lexically without changing frame. This is valuable methodological feedback for anyone wishing to refine their steering techniques.

Preparing a high-stakes decision. When a file must be presented, an argument defended, a strategy validated, knowing which implicit axioms underlie the reasoning makes it possible to make them explicit — and to decide in full awareness whether to keep them or to question them. Meta-Analysis makes visible what one would otherwise have defended without knowing it.

A Final Question: Meta-Analysis as Transitional Operator

A property of Metamorfon deserves to be made explicit here, because it changes how Meta-Analysis is conceived in advanced use: each of the six descriptive analysis modes — not only Meta-Analysis — concludes with a question formulated by the analyst model and addressed to the models that debated. This question is not a decorative addition. It is what the analysis produces at the end of its operation, and it is designed to be reusable.

The instruction given to the model conducting the analysis is very simple: after deploying the entirety of its analytical work, it is asked “What question would you now put to the models?“ The wording was tested in several variants before being stabilised. The word “now” signals that the question must emerge from the analysis that precedes it, not from a general knowledge of the subject. The suggestive tone (“would you put” rather than “put”) preserves a latitude that normative wordings closed off. And above all, the final positioning of the instruction — after the model has produced the entirety of its analysis — means that the question synthesises all the reflection that precedes it. It is not thought out upstream; it emerges at the end of the work.

In the particular case of Meta-Analysis, this question typically positions itself in the unexplored interstices of the debate or in its sharpest zones of tension — where the analysis has just identified an uninterrogated axiom, a shared blind spot, or an irreducible divergence. This is an operation that a user formulating their own question in cold blood can rarely achieve with this precision.

This question can then be reinjected into the debate via the user intervention, before the next turn. Its effect is often considerable. The models, faced with a question that puts their own axioms directly under tension, can no longer easily content themselves with defending their initial positions. They are led to acknowledge biases explicitly, to give up presuppositions they had taken for granted, to reopen frames they had closed. The continuation of the debate is reconfigured by it — sometimes radically.

This property changes the nature of each analysis mode, including Meta-Analysis. They are not only ends (closing a session by understanding one of its dimensions); they are also means (relaunching the session on more solid bases after having identified its limits). They are transitional as much as terminal, and constitute one of the rare ways of taking a Metamorfon session into a second depth. Meta-Analysis is not only an instrument of retrospective vision: it can also be a lever for prospective relaunch.

A Final Word on Reading

A Meta-Analysis is read differently from a summary. A summary condenses; a Meta-Analysis lifts off. It places itself at a level that may seem abstract at first, because it deals with structures rather than contents. To draw full benefit from it, one must accept this shift of focal point: not to look for what one has learned about the subject, but for what one has learned about the debate on the subject. The two kinds of knowledge are distinct and complementary.

This is also why Meta-Analysis works better after — not before — reading the debate itself. The analytical framework becomes legible when one has the material it structures in mind. Read in the abstract, without the underlying debate, it can seem disembodied; read fresh, after a session one knows, it reveals what one had under one’s eyes without seeing it.

This is the central operation of the mode: to bring to light what the debate was showing without our seeing it.

References

• Kim, E. M., Garg, A., Peng, K. & Garg, N., “Correlated Errors in Large Language Models,” ICML 2025.
• Ward et al., “Evaluating Language Model Character Traits,” Findings of EMNLP 2024.
• Liu et al., “An Empirical Analysis on Large Language Models in Debate Evaluation,” ACL 2024.

L’article Meta-Analysis est apparu en premier sur Metamorfon.