Verdicts and evidence

The kernel evaluates structured lore deterministically and offline. A caller may use a model outside the kernel to propose or interpret lore, but the kernel does not call a model or treat generated prose as a verdict. Attribution, review, and permission to rule belong to the consuming application.

The grade lattice

A verdict carries one of four verification grades:

checked > tested > attested > asserted

• checked: a procedure re-derives exactly from committed inputs.
• tested: a procedure has statistical evidence from a reproducible test.
• attested: a stored model judgment can be inspected and contested, but is not mechanically re-derived.
• asserted: a claim is taken as given, without a verifying procedure.

Support and grade are separate. Support records the decisions and assumptions that make a conclusion hold; grade describes the procedure used to reach the verdict. A deterministic proof check over authorially ruled premises can be checked even though the rulings themselves are asserted. A required procedure chain rolls up to its weakest grade; independent alternative procedures roll up to their strongest available grade. Neither rule turns repeated attestations into a test.

Attestations

Projects can contain committed attestation evidence. The project-backed dli.check verifies and replays the evidence against configured questions and reports current, missing, stale, or negative answers. Records retain the question, answer, model and prompt identifiers, sampling policy, rationale hash, fingerprint, and lineage needed to inspect them. A missing or stale answer is a finding, not a guessed answer.

Contract 2.0.0 does not provide an attestation-sweep write operation. The kernel reads existing evidence and reports demand; a consuming application owns any external evidence workflow and review policy. A model judgment is not proof of a claim and cannot silently create canon.

Structured claims

The engine checks the claims, decisions, and rules a caller supplies. It can report a contradiction, explain support and defeaters, or leave an unresolved claim possible. It cannot establish that a transcription faithfully captures a manuscript or that a model-generated assertion is true. Applications should inspect source material and apply their own review rules when that distinction matters.