The trust substrate
Everyone else trained the answers. We trained the reasoning.
The frontier labs built models that know an enormous amount. Eve-Genesis™ is the corpus that taught a model something different — how an expert in a field actually reasons. It is the asset the rest of the architecture is built to protect, and the one thing in the stack we do not rent.
“We’re not trying to know more than the frontier labs. We’re trying to reason better, in a way that travels anywhere.”Bill Faruki
The dataset started as riddles.
- 01
It began as riddles
Eve-Genesis began as a riddle dataset. The intuition was that a riddle isolates pure reasoning from domain knowledge — a riddle does not test what you know; it tests how you decompose, how you trace implications, how you resolve paradox. Training a Small Reasoning Model on riddles teaches it reasoning style independent of content.
- 02
A name for the logic underneath
The discovery came when the founder’s daughter, an undergraduate student, looked at the dataset and named the formal logic underneath it: deductive, inductive, abductive. The dataset, designed under another name, was already structured around the formal categories of reasoning recognised in philosophy.
- 03
Not conclusions — transitions
The recognition deepened: the dataset was doing things equivalent to dialectical reasoning, phenomenology, semiotics, hermeneutics, the Socratic method. It was not training the model on conclusions — on what answer goes with what input. It was training the model on conceptual transitions — on how to move between ideas.
- 04
The architectural posture
That is the architectural posture Eve-Genesis adopted explicitly: each record carries not just an answer but the cognitive operation that produces it — the reasoning mode used, the abstraction levels traversed, the alternative interpretations rejected. Trained on a corpus structured this way, the reasoner does not just know more facts. Its epistemic priors are shaped. It thinks differently.
That is what we mean by eve-genesis trains conceptual transitions, not conclusions.
Eight reasoning modes.
Every Eve-Genesis record is annotated with the cognitive operation that produces it. These are the eight modes the corpus is structured around — the formal categories of reasoning recognised in philosophy.
Deductive
Reasoning from general rule to certain conclusion. Truth-preserving when premises hold.
Inductive
Reasoning from examples and patterns to probable generalisation. Probabilistic, not certain.
Abductive
Inference to the best explanation. Central to diagnosis, intake, and scientific reasoning.
Analogical
Mapping one conceptual structure onto another. The substrate of case-based reasoning and teaching by example.
Dialectical
Concepts evolving through tension and resolution. Native to jurisprudence and advocacy.
Hermeneutic
Interpreting meaning contextually. Foundational to textual and source-based disciplines.
Phenomenological
Analysing how concepts appear in experience. Useful for instruction grounded in lived practice.
Socratic
Reasoning by question, by counterexample, by dialogue. The teaching posture of inquiry.
Four editions, four cognitive fingerprints.
Each edition shares the methodology but emphasises a different set of modes. The connected signature is the edition's fingerprint — and no two are alike.
Clinical Edition
In productionChironAI™ · Chiron
Emphasises Abductive · Analogical.
Education Edition
In productionArthurAI™ · Arthur
Emphasises Analogical · Socratic · Phenomenological.
Uṣūl Edition
In productionTheoAI™ · Theo
Emphasises Dialectical · Hermeneutic.
Law Edition
In productionJustineAI™ · Justine
Emphasises Analogical · Abductive · Dialectical.
100% synthetic by construction. No customer conversation is ever in the training set.
100% synthetic, by construction
No customer conversation, document, transcript, or interaction is ever in the training set. Not because of policy; because the architecture genuinely does not require it.
Per-domain editions
Each product’s reasoner is trained on its own Eve-Genesis edition. Knowledge in one domain does not leak into the cognitive posture of another.
Versioned and provenance-traceable
Every Eve-Genesis edition is versioned. Every record traces to a generation pass, with the reasoning structure documented at authoring time.
Carries no jurisdiction’s defaults
The trained reasoner is population-neutral by construction — it holds the shape of expert reasoning and none of a place’s specific rules. The governing law, the clinical guideline, the school of thought are supplied at runtime by the Digital Employee, in plain language you can read. The reasoning travels; the localization stays local, and stays visible.
Frontier-independent, and compounding
Frontier models are commodity consultants in the architecture; the reasoning is ours. And unlike a model snapshot, a reasoning corpus compounds — every edition deepens the methodology the next one starts from. The asset appreciates as the rented pieces get cheaper.
Ten editions. One per domain.
- Clinical EditionLiveSynthetic clinical reasoning corpus. Trains Eve-Healthcare.
- Education EditionLiveSynthetic pedagogical reasoning corpus. Trains Eve-Education.
- Legal EditionLiveSynthetic legal reasoning corpus. Trains Eve-Legal.
- Usul EditionLiveSynthetic Islamic jurisprudence reasoning corpus, grounded in usul al-fiqh. Trains Eve-Theology.
- Financial EditionRoadmapSynthetic financial reasoning corpus. Will train Eve-Finance.
- Insurance EditionRoadmapSynthetic insurance reasoning corpus. Will train Eve-Insurance.
- Real Estate EditionRoadmapSynthetic real estate reasoning corpus. Will train Eve-RealEstate.
- Commerce EditionRoadmapSynthetic retail and commerce reasoning corpus. Will train Eve-Retail.
- Marketing EditionRoadmapSynthetic marketing reasoning corpus. Will train Eve-Marketing.
- Engineering EditionRoadmapSynthetic technology and engineering reasoning corpus. Will train Eve-Technology.
The models will keep changing. The reasoning is ours.
Bias is not a stain you can scrub out of a single model — it is dissolved into the same weights that carry the model’s competence. Separating the reasoning from the knowledge from the jurisdiction is what makes it locatable instead. Why you can’t train bias out of a model.