Apodicta and the category
Retrieval changes what the model reads. It does not bind what the model writes.
Retrieval-augmented generation is the most common answer to hallucination, and it is a real improvement: the model drafts with the right documents in front of it. The step it does not change is the last one, where the sentence that cites the document is still produced by prediction.
Retrieval · and what Apodicta does instead
- The category Searches a corpus first and puts the retrieved passages into the prompt, so the model drafts with documents in front of it.
- Apodicta Derives the structure of the problem from the governing authority itself, then reasons inside that structure and proves each step to a source.GATED
- The category Improves what the model reads. The sentence that cites the document is still generated afterwards, by prediction.
- Apodicta Gates the finished draft. A line that cannot be tied to the passage it cites does not render at all.GATED
- The category A retrieved passage sitting next to a claim looks like support. Whether it carries the proposition is left to the reader.
- Apodicta Asks both questions of every authority: does it exist, and does it hold what the sentence attached to it says.GATED
- The category Failure is silent. A wrong answer arrives in the same confident typeface as a right one.
- Apodicta Failure is visible. What cannot be proved is held, stamped and named on the page, with the gap stated.GATED
- The category Quality moves with the index and the embedding model. Change either and the behaviour changes with it.
- Apodicta The rules are code outside the model, so swapping the model or the corpus does not move the guarantees.GATED
Where the category is the right choice
When the job is finding the material, retrieval is the tool. If you need the six documents in a 40,000-page production that mention a name, or the passage of a statute you half remember, a good retrieval system will get you there faster than anything else, and you will read the result yourself anyway. Apodicta uses retrieval too. The argument here is not that fetching is wrong. It is that fetching is not proving, and a filing needs the second one.
If nobody in your office is going to file what the system writes, and every output is a starting point a lawyer rewrites, the difference between the two approaches matters much less. It starts to matter the moment a sentence goes out with your name under it.
The longer versions
Why retrieval does not stop hallucination is the article with the sources. How it works is the architecture, gate by gate, andthe guardrails comparison covers the other common answer.