Say as little as possible
The common intuition holds that a good instruction tells the model what to find. Metamorfon rests on the opposite principle: an instruction should say as little as possible about the conclusion it expects. What it prescribes is not a result but a form — axes, an order of examination, a discipline of restitution. The content, it leaves to the debate.
The reason is simple. Any instruction that names what it hopes to find, finds it. A word carrying an implicit orientation pre-decides the analysis: asking to “spot the biases” manufactures biases; asking to “map what resists” does not presume that anything resists. The work of writing therefore consists in removing, one by one, the terms that incline. We call this lexical depopulation. Saint-Exupéry said of perfection that it is reached “non quand il n’y a plus rien à ajouter, mais quand il n’y a plus rien à retirer” — not when there is nothing left to add, but when there is nothing left to remove. An instruction obeys the same rule: it is ready when nothing more can be taken from it without making it inoperative.
Suggest, don’t command
This restraint also governs how a dialogue instruction addresses the models. It suggests more than it commands: it proposes a working posture rather than dictating each step, and deliberately leaves certain modalities implicit. This is not a slackening, but the continuation of the same principle. An imperative directive closes interpretation; a suggestion leaves latitude — the latitude a living exchange needs in order to produce something other than what was expected. Over-specifying the manner of debating would smother the antagonism that gives a session its value. And what can be left unsaid lightens the load: an accumulation of directives over-determines the model and scatters the attention the debate demands.
This way of writing was not decided at the outset; it was learned. Over several months of test sessions, the dialogue instructions were progressively lightened — made less imperative, more generic, stripped of the directives that dictated the form of an exchange. As they effaced themselves, the debates grew more fluid. The models stopped responding according to a pattern imposed from outside; relieved of part of that constraint, they turned more toward the substance of the argument than its formatting, and their voices grew more distinct — each recovering a style of its own that overly tight directives had tended to flatten. We do not claim to isolate a single cause, the models themselves having changed over the period; but the tendency recurred regularly enough, from one version to the next, to be trusted.
Analysis instructions are of another order. Restituting a debate reliably does not call for the same latitude as conducting one, and some of these instructions are more demanding — their function requires it. But whether it suggests or holds a form, no instruction ever touches the conclusion: neither the dialogue nor the analysis receives it in advance.
Equity and symmetry
Lexical depopulation settles a first plane: what the instruction says. By removing the terms that tilt, it prevents a conclusion from being whispered in advance. A second plane remains, more discreet: what each model receives. A neutral instruction, unevenly distributed, would cease to be neutral.
The architecture of cross dialogues and trilogues provides for this. At every turn, it supplies the debaters with the same context and the same instructions — none opens with an advantage, none speaks from a privileged position. Depopulation ensures that no conclusion is steered toward; the architecture, that no model is. One neutralizes the substance, the other the treatment.
From this follows something the sessions make legible. When a debate reveals an asymmetry — one model holding the initiative, setting the questions, pre-scripting its opponent while the other responds — it becomes legible precisely because the context does not vary. Under symmetric treatment and identical instructions, the observed divergence is not attributable to any inequality of treatment; it lies with the models and with the way each converts a signal that was nonetheless transmitted identically to all.
The generic endures, the prescriptive ages
An instruction cut to fit a precise model takes on its flaws and its era. It works for the span of one version, then degrades silently when the model changes — which happens fast. Simondon distinguished the abstract technical object, over-adapted to a particular milieu and made fragile the moment that milieu shifts, from the concrete object, more integrated and more open. An instruction over-fitted to a model is an abstract object in this sense: optimized for one state of the system, obsolete when it evolves. An instruction founded on principles rather than directives survives these transitions, because it describes a task and not a way of solving it. We write for the model we do not yet have.
It is also a matter of independence. The system sets competing models against one another, from different providers, by conviction as much as by method: epistemic diversity is a protection against the uniformity of a single source. An instruction that assumes a particular model betrays that commitment. A generic instruction honors it.
Transparency
The less the instruction and the user are present in the result, the more the result belongs to the debate. Intervention exists — it can tip a discussion from one register into another — but it is a lever pulled rarely, not a wheel held continuously. The value of a session lies in what the models produced among themselves, not in what was whispered to them.
From this follows a verification the system applies to itself. When the same debate is analyzed by two independent models, they draw the same map of its tensions: the same list of what resists, the same attribution of positions, the same transversal lines of fracture. They may gauge its depth differently — the structure is stable, its measure varies. But this convergence on the structure is the best proof that the map was in the material, not in the instruction. The role of an instruction is to be transparent enough for the material to show through.
The exceptions
The analysis modes describe; they do not judge. Two modes are exceptions, and in opposite ways. Argumentative evaluation is normative by construction — but even there, the principle holds: the instruction supplies criteria, not verdicts. It states by what to evaluate, never what to conclude. Source Verification, by contrast, does conclude: it returns verdicts — confirmed, partially correct, not found. But it returns them only on facts verifiable against an external source, never on the value of the ideas or the soundness of the positions. The one states by what to judge a line of reasoning; the other settles what is attestable outside the debate. In both cases, the boundary between describing and judging remains explicit — and it is precisely because it is explicit that it can be crossed without confusion.
A successful instruction is the one the reader, seeing the result, would not have guessed was there. It does not show itself in what it produces. It has done its work when it can no longer be seen.