Skip to content

Concepts

This section is about the structural ideas behind Eigenius, written for the reader who wants to understand the framework without committing to learning the platform. If you want to try the platform, head to Docs instead.

The short video below frames the core motivation: AI reasoning agents are dangerously unreliable today not because they make mistakes (so do humans), but because their reasoning is opaque prose that cannot be mechanically inspected. Eigenius proposes a typed substrate that turns the AI’s reasoning into a typed chain commit the kernel can verify — closing the feedback loop that AI coding agents already have but reasoning agents lack.

Why are AI coding agents incredibly effective, while AI reasoning agents remain dangerously unreliable? The answer is the same one that makes compilers useful: a type discipline the agent must produce against.

The four warrant categories

Every claim in an Eigenius chain lands in exactly one of four categories of warrant. Three are mechanical — anyone with read access can re-execute the warrant; the fourth is authoritative — the declarer’s identity and citation chain are examinable. The distinction is what an AI agent must declare when it commits a claim, and what stops a hallucination from masquerading as a mechanical derivation.

Mechanical warrants re-executable
Observed Instrumental

Raw measured data with unbroken instrument provenance.

Audit fingerprint Instrument run, time, operator.

Example: Six raw IC₅₀ readings from a Kinase-Glo plate run.

Derived Computational

Outputs of deterministic computation, rerunnable byte-identically.

Audit fingerprint Procedure, inputs, kernel that ran it.

Example: A t-statistic and p-value the statistics institution emitted.

Verified Formal-proof

Claim accepted by a machine-checked mathematical proof.

Audit fingerprint Proof artifact, checker, independent re-execution.

Example: A Lean-verified regulatory primary-endpoint result.

Authoritative warrant examinable
Declared Authoritative

Asserted by a named human, organisation, or policy document.

Audit fingerprint Declarer identity, citation, organisational authority.

Example: A 100 nM threshold cited from a methodology document.

  • Domain bridges — the declarative first move in crossing between domains: formulate the entities and propositions of a specific domain and state the implications that link them to a neighbouring vocabulary, as a citable, refutable, chain-resident artifact.
  • Institutions — the modular dispatch architecture. How institutions lift domain bridges from flat data into live computation and reasoning, and how comorphisms mediate between runtimes at typed seams.
  • Justification logic (coming soon) — the modal calculus that underlies the four-warrant taxonomy. For the academically curious.

When you’re ready to see all of this end-to-end in a real workflow, the drug-screening example walks the chain from raw IC₅₀ plate readings to the final designation.