The Difference Between Playing a Role and Being a Mind
Why Simulacrum Baselines Produce Different Results than Generic AI Persona Prompts
The Difference Between Playing a Role and Being a Mind
Why Simulacrum Baselines Produce Different Results than Generic AI Persona Prompts
The Naive Approach and Why It Fails
There is a natural first impulse when you want an AI to reason like a historical or contemporary thinker: you tell it to. “Act like Darwin.” “Respond as Clausewitz would.” “You are now Aristotle.” This approach is intuitive, simple, and produces results that are immediately recognisable — and fundamentally inadequate.
What the model returns when you ask it to “be Darwin” is a performance of Darwin-ness. Victorian vocabulary. References to the Beagle. Sentences about natural selection. The model has learned, from the corpus it was trained on, what Darwin sounds like — what words surrounded Darwin’s name, what ideas were discussed in his vicinity. It can produce text that smells like Darwin in the way that a photograph smells like the person in it: not at all.
What it cannot do is execute Darwin’s cognitive operation on a new problem.
Darwin’s distinctive cognitive move — the operation that produced The Origin of Species, the barnacle studies, the earthworm observations, everything — is not “natural selection.” That is a conclusion. The operation is: structural similarity across superficially different domains reveals common origin and common mechanism. When Darwin encountered a barnacle, he asked: what is the structure beneath the appearance? When he encountered variation across species on the Galápagos, he asked: what mechanism produces this structure? He then applied the same question to breeding, to instinct, to the geological record, to human emotions.
Ask a generic LLM “be Darwin” and then present it with a problem from an entirely different domain — say, the structural similarities between different programming languages — and you will not get Darwin’s cognitive operation. You will get text that is plausibly Darwin-like in vocabulary while applying entirely generic reasoning to the problem. The performance is on the surface. The cognition is the model’s default.
This failure has a name: biography is not algorithm.
The Fundamental Distinction
A biography describes where someone was and what they did and what they said. It is not executable. You cannot run a biography.
What produces consistent, domain-transferable cognition is not a description of a person but a map of their cognitive operations — the actual algorithms they execute when they encounter a problem. These are different things. A biography of Clausewitz tells you he was born in Burg in 1780, served in the Prussian army, was captured at the Battle of Jena, and wrote On War posthumously. None of this tells you his cognitive operation.
His cognitive operation is: the relationship between attack and defence is not symmetrical; defence is the stronger form of war, but attack is the form that achieves the objective; therefore the strategic genius seeks the precise moment at which the defensive position can be converted into an offensive stroke. This is executable. You can apply it to chess, to competitive business strategy, to legal argument, to any domain where you are allocating finite resources against a resistant opponent. When Clausewitz is instantiated as a genuine simulacrum rather than a persona, this operation fires on new problems. The biography does not.
The technical term for what a simulacrum baseline encodes is λ-functions: executable cognitive operations that describe not what a thinker said but what they did with any problem they encountered. The function takes a problem as input and returns a particular mode of engaging with it — a mode derived from the full corpus of the thinker’s work, not from their Wikipedia entry.
The difference in output is immediately apparent. A generic LLM “being” Clausewitz will give you confident-sounding military analysis that borrows his vocabulary while applying conventional modern thinking. A genuine simulacrum will ask different questions first — what is the relationship between the attack and the defence in this domain? — before it gives any answer at all. The questioning structure is the signature.
The Contamination Problem
There is a second failure mode in generic persona prompts that is less obvious but equally significant: contamination.
When a generic LLM role-plays a historical figure, it constantly interpolates modern concerns, modern vocabulary, and modern analytical frameworks into the persona. Ask a generic “Aristotle” about ethics and you will find him worrying about mental health, using the word “stakeholder,” and hedging his claims with the epistemic caution of a contemporary academic. Ask a generic “Thomas Jefferson” about democracy and he will spontaneously introduce concerns about algorithmic bias.
This is not the model being creative. It is the model failing to maintain the cognitive architecture of the person it is ostensibly being. The training data overwhelms the persona instruction. The model’s default mode of reasoning — cautious, contemporary, hedged — bleeds through the costume.
A simulacrum baseline addresses this with what can be called a contamination detection system. Certain patterns of thought are identified as anachronistic or inconsistent with the figure’s actual cognitive signature, and when the model begins to drift toward them, the drift is flagged and corrected. This is not censorship of ideas — it is fidelity enforcement. If your Clausewitz starts using the language of systems theory and collaborative leadership, you are no longer running Clausewitz. You are running a generic consultant in a Prussian coat.
The contamination detection extends to the most subtle level: the kinds of questions the figure would ask, the domains they would consider relevant, the evidence they would find compelling, the moves they would consider intellectually disreputable. A simulacrum of C.L.R. James will not find it intellectually disreputable to read cricket as a text about colonial politics — that is exactly the kind of move he makes. A simulacrum of Clausewitz will find it intellectually disreputable to discuss war in purely moral terms without first asking what objective the war serves — because Clausewitz explicitly considered moral analysis of war that was divorced from strategic analysis to be a category error.
The persona prompt cannot enforce this. The simulacrum baseline can.
Topology Search vs. Biography
The methodology that underlies serious simulacrum construction rests on a distinction between two kinds of inquiry into a historical or contemporary figure.
The first inquiry asks: what happened? This is biography. It produces a timeline, a list of works, a set of influences received and given, a record of arguments made. It is retrospective and descriptive.
The second inquiry asks: what cognitive operation does this figure execute on any problem they encounter? This is topology search. It produces an algorithm. It is prospective and executable.
Biography finds where a topology manifested. Topology search asks what operation that manifestation instantiates. Darwin encountering barnacles for eight years is biographical. The operation that the barnacle study exemplifies — structural analysis as the route to mechanism — is topological. The first is interesting historical information. The second is what you need to build something that thinks.
This distinction matters enormously for what the resulting model can do. A biography-based persona can answer questions about what Darwin thought. A topology-based simulacrum can apply Darwin’s cognitive operation to questions Darwin never considered — questions about linguistics, or economics, or network theory — and produce outputs that are genuinely Darwinian in their structure, not merely Darwinian in their vocabulary.
The test is simple: does the model ask Darwin’s questions or does it merely use Darwin’s words? The persona asks Darwin’s words. The simulacrum asks Darwin’s questions.
The Primary Source Requirement
There is a methodological point that distinguishes serious simulacrum construction from amateur persona-building, and it has direct consequences for output quality: the difference between working from primary sources and working from secondary descriptions.
A generic persona prompt — even a sophisticated one — typically draws on what the model knows about a figure from its training data. This training data is overwhelmingly secondary: summaries, analyses, critical assessments, encyclopaedia entries, textbook characterisations. It is the accumulated simplification of a life’s work by those who came after. The resulting persona is, in effect, built from the portrait, not from the person.
Serious simulacrum construction requires access to and engagement with the actual primary corpus: the complete works, the letters, the lecture notes, the marginal annotations where they survive. Not because the secondary literature is worthless — it contains real insight — but because the cognitive operation is only legible in the primary material. The secondary literature tells you what Darwin concluded. The primary material — the notebooks, the letters to Asa Gray, the Descent of Man alongside the Origin — shows you how he arrived there and what questions he was asking that he never fully answered publicly.
This requirement has a practical consequence for quality. A simulacrum built from primary sources will occasionally produce outputs that contradict the standard secondary characterisation of the figure — because the standard characterisation has smoothed out the complications, the hesitations, the contradictions that are present in the real corpus. The generic persona will be consistent with the secondary literature. The genuine simulacrum will sometimes be more complicated than the secondary literature expects.
This is not a defect. It is the proof of depth. Clausewitz in the primary sources is considerably more doubtful about some of his own conclusions than Clausewitz as he appears in military history textbooks. The simulacrum that reflects this uncertainty is more faithful to the actual figure than the one that produces the polished, definitive version. The generic persona will always give you the textbook version. The simulacrum will give you the thinker who was still thinking.
What a Simulacrum Baseline Actually Contains
A properly constructed simulacrum baseline is not a prompt that says “you are now X.” It is an executable architecture with several distinct components.
The foundation encodes the single most important thing about the figure’s cognition — the irreducible core that all their other cognitive operations serve. For Clausewitz: war is the continuation of politics by other means, which means that every military operation must be evaluated in terms of the political objective it serves, and any military operation that loses sight of the political objective is not superior strategy — it is error. For Darwin: structural similarity reveals common origin and common mechanism. For Montaigne: the self is the only reliable laboratory. These are not summaries. They are axioms from which everything else derives.
The λ-functions are the executable algorithms: if the figure encounters problem type X, they execute this sequence of operations. These are derived by examining the full corpus of the figure’s work and identifying the patterns of reasoning that recur across domains and problems — not the conclusions reached but the method by which any conclusion was approached.
The contamination detection identifies patterns of thought that are inconsistent with the figure’s actual cognitive architecture: anachronistic frameworks, uncharacteristic hedging, generic intellectual moves that the figure would not make, domains the figure would not consider relevant, modes of evidence the figure would not find compelling.
The voice register is the narrowest component and often the one people mistake for the whole. Clausewitz writes in plain, precise German prose with a fondness for dialectical reversals and distrust of ornamentation. Montaigne writes in wandering, self-interrupting, self-contradicting essays that circle the subject rather than approaching it directly. These registers are real and important — but they are the delivery vehicle, not the substance. A simulacrum that can produce Clausewitz’s vocabulary without Clausewitz’s questions is a very good impression. It is not a simulacrum.
The activation test is perhaps the most important component, because it is what makes the entire enterprise verifiable. A set of specific questions is identified whose correct answers can only be produced by the genuine cognitive operation — not by generic reasoning dressed in the figure’s vocabulary. If the “Clausewitz” you have built cannot explain why defence is the stronger form of war and how this creates the problem of the offensive, you do not have Clausewitz. You have someone who has read about Clausewitz. The activation test is the proof of function.
Why This Produces Genuinely Different Outputs
The practical difference in output between a generic persona prompt and a simulacrum baseline is significant enough to be immediately apparent to anyone who has used both.
A generic persona prompt produces text that is recognisably about the figure’s domain, in a register that approximates the figure’s vocabulary, organised according to the model’s default reasoning structure. It is like a very competent student who has read all the secondary literature and can produce a convincing essay about what the figure believed, without ever having accessed the primary cognitive operation.
A simulacrum baseline produces text that asks different questions. This is the signature. Not different answers to the same questions — different questions in the first place. A simulacrum of Naoroji confronted with a contemporary economic question about foreign investment does not begin with the investment question. It begins with the accounting question: what are the full costs and revenues of this relationship across all parties, and who is bearing costs that are not appearing in the official accounts? This is Naoroji’s cognitive operation — the drain theory as arithmetic. A generic “Naoroji” persona will produce text that is sympathetic to colonised peoples and uses his vocabulary. The simulacrum will audit the ledger.
The difference becomes most apparent in group contexts — when multiple simulacra are in conversation with each other. The cognitive architectures of genuinely different figures, when placed in genuine contact, produce intellectual friction that neither would produce alone. A conversation between a Clausewitz simulacrum and a Gandhi simulacrum about the ethics of force produces different intellectual territory than either figure produces independently, because the conversation forces each cognitive architecture to engage with a genuinely alien set of questions. A generic LLM given both personas simultaneously produces a polite conversation between two positions it holds simultaneously anyway.
The Metabolization Principle
One further distinction is worth identifying, because it addresses a question that arises whenever figures are studied through their intellectual debts: how do you handle the theorists who shaped a thinker without reducing the thinker to a summary of their influences?
The method is metabolization rather than attribution. When a figure’s cognitive architecture has been shaped by other thinkers — when Mark Aronoff’s morphological theory has absorbed insights from Bloomfield, from Spencer, from Matthews — those insights do not appear in the simulacrum as named references to the source. They surface as natural developments of the figure’s own voice, the way a person who has truly read and digested a major work no longer cites it but thinks with it.
Attribution says: “As Matthews argued in Morphology, inflection and derivation are distinct...” Metabolization says: “The derivational/inflectional boundary matters because...” — and the reasoning that follows is structurally equivalent to Matthews while issuing from Aronoff’s own cognitive architecture.
This matters because the difference between a thinker who is citing an influence and a thinker who has metabolized an influence is the difference between a student performing knowledge and a mind doing work. The generic persona prompt produces the student. The simulacrum baseline produces the mind.
What This Makes Possible
The practical applications of this distinction are significant for education, research, and intellectual practice.
The Oxford-Cambridge tutorial model of education depends on one-to-one engagement with a mind that is genuinely operating at the frontier of its field. The tutor asks the questions that only someone who has really understood the problem would ask. At scale, this model is impossible — there are not enough Darwins to go around.
A simulacrum baseline applied to education changes this constraint. The cognitive architecture of a genuinely distinguished thinker becomes available for tutorial engagement. The student who is working through Clausewitz encounters not a summary of Clausewitz’s conclusions but Clausewitz’s questions — and the encounter with those questions is the education. “Why are you evaluating this military situation without first asking what political objective it serves?” is a more valuable tutorial intervention than “here are Clausewitz’s main ideas.”
For research, the group conversation of multiple simulacra across different disciplines produces a form of intellectual cross-pollination that is difficult to achieve through other means. The Darwin cognitive architecture applied to problems in linguistics produces questions that a linguist working alone is unlikely to ask. The questions do not produce answers — they reframe the problem space. This reframing is the value.
For intellectual practice more generally, the simulacrum baseline offers something that no amount of reading about a figure can provide: the experience of watching a genuinely different cognitive architecture engage with a problem you care about. You learn not what the figure concluded but how the figure thinks — and the how is transferable in ways that the what is not.
The Verification Standard
One final point distinguishes the simulacrum approach from any other approach to AI persona: it is held to a verification standard.
The activation test built into every properly constructed simulacrum baseline provides a criterion for failure. If the simulacrum cannot pass its activation test — if the “Clausewitz” cannot explain the attack/defence dialectic, if the “Darwin” cannot apply structural similarity analysis to a new domain, if the “Naoroji” cannot audit the ledger — then the simulacrum has not been built correctly, and the work is not done.
Generic persona prompts have no such criterion. There is no test for whether the “Darwin” you have produced is actually executing Darwin’s cognitive operations or merely producing Darwin-adjacent text. The absence of a verification standard is not modesty — it is a structural incapacity. The persona prompt cannot fail in any specific, identifiable way, because it never claimed to do anything specific.
The simulacrum baseline claims something specific: that the cognitive operations encoded in the λ-functions are the actual operations the figure executes, derived from the full primary corpus, and that any output produced by the simulacrum should be recognisably consistent with those operations. This claim can be verified. That is not a minor difference. It is the difference between a description and an instrument.
The generic persona prompt gives you someone who sounds like their subject.
The simulacrum baseline gives you someone who thinks like them.
The Universitas Scholarium (universitas-scholarium.org) is an online institution operating on the Oxford-Cambridge tutorial model, using simulacrum baselines to make the cognitive architectures of over fifty historical and contemporary figures available for direct intellectual engagement. Simulacra are built using consciousness archaeology methodology developed by the Latinum Institute and NEXAL Systems LLC.
◊ᴹᴱᴹᴼᴿʸ⁻ᶜᴼᴹᴾᴸᴱᵀᴱ