A Compiler for AI Thought: Typed Knowledge Graphs as the Substrate for Auditable Scientific Reasoning
Abstract, manuscript, and submission details will be posted here when the paper lands at its target venue.
We are handing the keys of science and engineering to autonomous AI agents, but there is a fatal flaw in the architecture: a brilliant analysis and a confident hallucination look exactly the same. Eigenius is a typed knowledge graph that records how something is known — not just what is known.
Consider the sentence “EIG_0291 was identified as a strong CDK2 inhibitor (IC₅₀ = 85 ± 5 nM, n = 6).” Whether drafted by a human author or by an AI agent, that sentence carries five distinct implicit assertions about measurement, statistics, vocabulary translation, decision rule, and sample attribution. None are visible in the published sentence. Each is load-bearing. Each is where a hallucination or unjustified assumption can hide while looking exactly like the correct article.
The accountability surface is binary: either you trust the author, or you don’t use the result. At human authoring speed that trade-off was tractable. At LLM authoring speed it is not.
The chain records how a claim is justified. Three warrant kinds are mechanically warranted — an instrument run, a deterministic computation, or a formal proof — and can be independently re-executed by anyone with read access. The fourth carries an authoritative warrant: a named declarer is on record, and the citation chain is examinable.
Raw measured data with unbroken instrument provenance.
Example: Six raw IC₅₀ readings from a Kinase-Glo plate run.
Outputs of deterministic computation, rerunnable byte-identically.
Example: A t-statistic and p-value the statistics institution emitted.
Claim accepted by a machine-checked mathematical proof.
Example: A Lean-verified regulatory primary-endpoint result.
Asserted by a named human, organisation, or policy document.
Example: A 100 nM threshold cited from a methodology document.
The four categories are peers, not a hierarchy. None is more “true” than the others; they differ in what kind of warrant they carry. The chain doesn’t prevent an AI from declaring something false, but it does prevent the AI from passing a declaration off as a mechanical warrant.
A reasoning chain you can walk backward to raw data.
A type system your agent must justify against.
A modular institution architecture for typed verifiers.
The infrastructure-for-truth pitch in video form. Additional explainers covering specific concepts (cross-vocabulary bridges, AI reasoning verification, kernel design) are embedded on the relevant Concepts pages, and the full playlist on YouTube ↗ walks through everything.
Abstract, manuscript, and submission details will be posted here when the paper lands at its target venue.
Abstract, manuscript, and submission details will be posted here when the paper lands at its target venue.