Eve-Genesis is reasoning-style conditioning, not just fine-tuning.

There is a taxonomy of training interventions for large language models that most marketing copy in this market deliberately blurs. I want to draw the distinctions cleanly because the distinctions matter for understanding what Eve-Genesis actually does and where it sits.

Prompt engineering changes what the model is asked. A well-engineered prompt elicits better responses from a frozen model. Nothing about the model is altered. Almost every "AI for X" product on the market today is at this level, layering its system prompt on top of a frontier model.

Retrieval-Augmented Generation (RAG) changes what the model sees. The model is given relevant documents at inference time and asked to ground its response in them. Useful. Particularly so when factual accuracy matters. But: the model is still the same model. RAG makes the model read better. It does not make the model think differently.

Standard fine-tuning changes what the model knows. The weights are updated on a domain corpus. Facts that were not previously present are now associated. Style mimics the training corpus. This is the level most "domain AI" companies operate at, and it is meaningful, but it is bounded by what the underlying foundation model can be taught to know.

Reasoning-style conditioning changes how the model thinks. The training corpus is structured to carry not just instruction- response pairs but the cognitive operations that produce the responses — the reasoning mode, the abstraction levels traversed, the alternative interpretations considered. The model that emerges has been shaped at the epistemic level. It does not just know more. It reasons differently.

The depth of the intervention determines the moat. A prompt-engineered AI product has no moat — the system prompt is copyable in an afternoon. A RAG product has a thin moat — the document index can be rebuilt by anyone with access to the source material. A fine-tuned product has a meaningful moat — reproducing the training corpus and the training run is non-trivial.

A reasoning-style-conditioned product has a categorically different moat. The corpus structure is the hard part to reproduce. The reasoning-mode taxonomy had to be derived by working backward from a methodology that mostly lives in disciplines that do not speak machine-learning. Encoding the reasoning of a senior clinician, a senior educator, a senior scholar, or a senior attorney into a training corpus is not an engineering exercise. It is a working session with someone who can articulate their cognitive moves — a rare skill even among senior practitioners.

Each Eve-Genesis edition produces a Small Reasoning Model whose epistemic priors are shaped to match its discipline's actual reasoning practice. On Arthur, the F5/reasoner inherits analogical reasoning (the substrate of teaching by example), Socratic questioning (the inquiry posture), and phenomenological grounding (lived-experience-anchored abstraction). On Chiron, the F5/reasoner inherits abductive reasoning (the mode of differential diagnosis). On Theo, dialectical and hermeneutic (the mode of Uṣūl). On Justine, analogical, abductive, and dialectical (the mode of case-based legal practice).

That inheritance is what makes the reasoner an orchestrator instead of a commodity LLM. The reasoner does not retrieve documents and read them; the reasoner thinks about the problem in the discipline's native idiom. When the reasoner delegates a sub-problem to a frontier consultant, it knows what to ask — because it knows the discipline.

People sometimes ask whether reasoning-style conditioning becomes obsolete when frontier models get better. It does not. Frontier-model improvement helps every consultant in our architecture; it does not help our competitors' lack of an orchestrator. As GPT-6 launches, our reasoner gets a stronger consultant. Our competitors get the same. The orchestrator IP is unchanged. The architecture appreciates as the frontier moves.

How Eve-Genesis was invented →