Apodicta and the category
A guardrail scores a probability. A gate decides whether a line may print.
Guardrails, verifiers and output scorers all do the same useful thing: they look at what a model produced and judge it. The judgement is a likelihood, and the label is attached to a sentence that is already in the document. That is the gap this page is about.
Guardrails · and what Apodicta does instead
- The category Runs a classifier or a second model over the output and scores how likely it is to be wrong, unsafe or off policy.
- Apodicta Runs a gauntlet of executable rules over the output: every assertion must carry a cite, every inference a proof.GATED
- The category The score is a probability, so someone has to choose a threshold. That choice is a decision about how much wrongness to ship.
- Apodicta There is no threshold. A claim either resolves to the source it cites, or it does not render.GATED
- The category A flag leaves the sentence in the document, labelled. The document is still the one you are about to file.
- Apodicta A failed claim is pulled out of the draft, stamped, and shown with the reason it was held.GATED
- The category The checker is usually another model, so it can be confidently wrong in the same way as the thing it is checking.
- Apodicta The checks are deterministic code, so the same draft and the same sources produce the same verdict every time.GATED
- The category Behaviour is tuned per deployment, so the guarantee lives in configuration and moves when the model moves.
- Apodicta The enforcement boundary is code outside the model, so a local model and a frontier model are governed identically.GATED
Where the category is the right choice
Most of what a guardrail is built for has nothing to do with citations. Blocking prompt injection, keeping a customer-facing assistant on topic, catching abuse, holding a tone: those are jobs a scorer does well, and a hard gate would be the wrong shape for them, because there is no source to check a claim against. If the question is whether an output is appropriate rather than whether it is true, a probability is the honest answer.
A scorer is also cheap to add to a system you already have. A gate needs the sources in hand and a definition of what would count as proof, which is work. Where nothing is filed and nobody signs, that work may not be worth doing.
The longer versions
What a citation gate is sets out the two questions a gate has to ask. How it works shows the five gates in series, andfail-closed is the property that makes the difference above hold.