Evidence and unknowns

The vocabulary Orneos uses about proof — what a source reference means, what it does not, and why unknowns are stated rather than omitted.

Reference Updated 2026-09-22

This page is about language. It exists because most tooling in this space is imprecise about what it is claiming, and that imprecision is what makes the output untrustworthy.

Three levels, kept apart

When a review says something about a requirement, it is making one of three claims:

Supported. There is material here that speaks to this requirement, and it is named so you can open it.

Open question. Something specific is unresolved, and we can say what it is. “The criterion says recovery should be safe; the linked page says safe recovery was left open with no agreed behaviour” is an open question. It has a shape and someone could resolve it.

Unknown. We could not establish this either way. No evidence was supplied, or we could not see the relevant material.

Collapsing these is the failure mode. A tool that reports “3 of 5 criteria met” has collapsed all three into one number, and the number is not recoverable back into the thing a reviewer needs to act on.

A source reference is not proof

This is the single most important sentence in these docs, so it gets its own heading.

When a context bundle says a claim came from a particular Vault page, that means: you can open that page. It does not mean the page establishes the requirement.

The reference is a pointer, not an argument. Treating it as an argument is how a chain of plausible-looking references ends up supporting a conclusion nobody checked.

Unknowns are stated, not omitted

The tempting design is to report what you found and stay quiet about what you could not check. It reads better and it is much more dangerous, because a reader cannot distinguish “this was checked and is fine” from “this was not checked”.

So a gap is named. “No evidence was supplied for criterion R3” is a finding. It is often the finding.

The cost is honest: it makes output longer and occasionally makes us look less capable. We would rather look less capable than have you approve something on the strength of a silence.

A decision covers what you actually looked at

An acceptance is bound to the thing that was reviewed — a particular Intent revision and a particular artifact.

Accepting one principle in a review does not mean:

  • the implementation is correct;
  • the task is finished;
  • a later revision of the agreement is covered;
  • an unresolved question has been resolved.

If the agreement moves to a new revision, the earlier acceptance does not follow it. This is the same rule that governs Intent approval, applied to the decision rather than the agreement, and for the same reason: an approval that silently expands to cover things nobody looked at is not an approval.

Where AI sits in this

Where a model is involved — drafting an Intent, answering a question in Insights — that involvement is disclosed, and the output is a proposal, not a finding.

A person reads it and decides. A generated draft nobody read is not an agreement; a generated summary nobody checked is not evidence. See AI in Orneos.

Something here wrong, missing or out of date? Tell us at support@orneos.com — corrections to these pages are welcome and we would rather hear it than have you work around it.