Deterministic grounding

Deterministic grounding is our name for grounding built so that the same question over the same corpus returns the same evidence every time, and so that the properties that make an answer trustworthy are enforced before a release ships rather than measured after it does. We are naming it because the thing it describes does not have a name, and the category it belongs to is currently sorted by technique rather than by guarantee.

What does “deterministic” actually claim here?

It claims reproducibility, not omniscience. Ask the same question of the same corpus and the grounding service returns the same evidence, in the same order, with the same citations. Run it a hundred times and the output hashes match. That is a narrow claim and a testable one, which is why it is the claim worth making.

It is not a claim that the answer is always right, and it is not a claim about the language model sitting on top. A model still phrases the answer, and phrasing varies. What stops varying is the evidence underneath it. Deterministic grounding moves the part that has to be right out of the probabilistic layer and into a layer you can put a probe on.

Why is reproducibility the property that matters?

Because it is the one an auditor can use. A system that is right most of the time and different every time cannot be checked. If an engineer signs off on a procedure in March and someone asks in November what the system said and why, a probabilistic retrieval layer cannot answer, because it no longer produces the same retrieval. A deterministic one can be re-run.

This is the same reason a calculator is trusted and a guess is not. Not because the calculator is powerful, but because the path from input to output is the same path every time and every step on it is accountable. Reliability that cannot be reproduced is a run of good luck that nobody has measured the end of yet.

How is this different from RAG?

Retrieval-augmented generation is a family of techniques. Deterministic grounding is a statement about what the layer guarantees. The distinction matters because it changes what you evaluate.

Ordinary RAG chunks documents, embeds the pieces, and returns the few that score closest to the question. It is quick to stand up and its reliability is a function of how your particular documents happen to chunk, which is why it works in a demo and degrades in production. The advanced variants, GraphRAG, corrective RAG, agentic retrieval, are real improvements and they work. They buy reliability with complexity and latency: more moving parts, more tuning surface, more round trips before an answer comes back.

Deterministic grounding takes the cost at ingestion instead. The structure of a document, its tables, its cross references, and the exceptions that qualify its claims are preserved when the document goes in, so retrieval is a lookup against a structure rather than a similarity search against fragments. That removes the tuning surface, which is where most ordinary retrieval failures actually live, and it keeps latency low at query time because the expensive work already happened.

What does it guarantee?

Guarantees are only worth the enforcement behind them, so these are release gates. Each one blocks a release if it fails.

• Losslessness. Nothing in an admitted document is silently dropped between ingestion and retrieval.
• Caveat attachment. A claim is not returned without the exceptions and conditions that qualify it.
• Citation validity. Every claim resolves to the passage and revision it came from.
• Determinism. Identical output across one hundred repeats of the same query.
• Boundaries. No answer is returned from outside the declared corpus.
• Qualified absence. A question the corpus cannot answer is answered “not found”, with what would be needed to answer it, never with an invention.

The honest form of the claim is that the structural failures are removed by construction and the operational ones are made visible rather than silent. A system that cannot answer a question is not a failure. A system that cannot tell you it could not answer is.

Is this a category or a marketing word?

A fair question, and the answer is that it becomes a category if it survives being tested, not because we said it. So the useful version of this page is the one that tells you how to falsify it. Bring a corpus where a clause and its exception sit pages apart. Ask the question that depends on both. Run it a hundred times. Check whether the evidence is identical each time, whether the exception came back attached, and whether the system told you what it did not read.

That test is what a grounding teardown is, and we would rather you ran it than took the term on trust.

Related terms

• Grounding: the discipline this is a specific, enforced form of.
• Traceability: the property that makes a deterministic answer checkable.
• Silent omission: the failure the boundaries and absence gates exist to remove.
• DSAI: what grounding produces once a domain is captured.