Retrieval-augmented generation was supposed to fix hallucination. Give the model the facts and it will stop inventing them. What the industry learned instead is that the model faithfully reports whatever the store hands it, including the store's own uncertainty, which the store never mentioned. A record with one weak source reads exactly like a record with forty. A field that is genuinely unknown reads exactly like a field that was redacted, or never collected, or is above the reader's clearance. Two sources that disagree have already been collapsed to one value before the model sees them. The model then speaks with a confidence the data never had, and we call it hallucination.
The database is part of the model. Its silences and its flattenings become the model's assertions. This article is for the people building retrieval, agent memory and decision systems over consequential data, and for the leaders being asked why the system was so sure. It describes what a store has to carry for the model on top of it to be honest, and what 8DB does about each.
Confidence is a column, not a footnote
In 8DB every record carries a multi-axis confidence vector as a first-class storage column. The axes answer the questions a careful analyst asks: is this sourced or assumed, how many independent anchors support it, are those sources truly independent or do they share an upstream, how recent is it, how good was the method that produced it, how many authorities agree, and does the source have authority over the claim. A retrieval system reading 8DB does not receive "the answer". It receives the answer and the shape of the evidence behind it, and it can be built to say so.1
This is different from a relevance score. Relevance says the record matches the query. Confidence says how much the record should be believed, and why. A model given both can distinguish "I found it" from "I know it".
Eight kinds of nothing
Null is the most dangerous value in a knowledge base, because it means everything and the model will pick one. 8DB distinguishes eight kinds of absence rather than one null: absent, unknown, redacted, existence-hidden, below clearance, stale, frozen and sealed. A record with no birth date because no source recorded one is not the same as a record whose birth date is redacted, or held above the reader's clearance, or last observed too long ago to trust, and the store keeps those distinct. Each of those is a different true statement, and a model that receives the distinction can make the right one instead of a guess.
The store also refuses to answer beyond its confidence frontier. A question about a field the store cannot support returns the explicit gap, stated, rather than a value. This is the structural opposite of a language model's default behaviour, and it is the property that lets a system built on 8DB say "unknown" and mean it.
Conflicts are evidence, not noise
When sources disagree, most pipelines pick a winner before the model sees the record, and the disagreement disappears. 8DB keeps every value, tags the conflict, records which attributes are unanimous and which are contested, and surfaces the item for review. On a public sanctions corpus of roughly 4.28 million records, one individual appears with six birth years across authoritative lists and another with three nationalities, and the store held all of them with their sources rather than averaging them away.2 A model that sees the conflict can report it. A model that does not will assert one side.
Provenance you can query
Every record in 8DB is content-addressed and attributable to its source, every merge or derivation carries a provenance chain with a signed, reversible anchor, and datasets are authenticated with one ML-DSA-87 signature over a SHA-384 Merkle root with a compact inclusion proof per record. "Why did the system say this" becomes a query that returns the source atoms, their confidence, and the governance state at the time, not a forensic reconstruction over logs. A reviewer can be handed one record, its proof and the signed root and verify it without trusting the operator.
What all of this changes is the honesty of the inputs, which is the part the model cannot fix for itself.3
What to bring us
- One question your system gets wrong. A case where the model was confidently wrong, or confidently silent. We will show what the store handed it and what a confidence-carrying store would have handed it instead.
- Your retrieval pipeline. Where it flattens, where it picks winners, where null means five things. That becomes the acceptance plan.
- A corpus slice. We will load it and return every record with its confidence vector, its provenance and its gaps stated.
Write to hello@8braid.com with the subject "Evidence-carrying store for retrieval". You will hear back from an engineer.
Sources and further reading
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- NIST AI Risk Management Framework
- 8DB: Entity Resolution That Shows Its Work
- 8DB: An Agent Can Only Leak What Its Process Can Decrypt
Footnotes
-
The confidence, absence and provenance mechanisms are in the engine and are exercised in our independent evaluation package as supporting demonstrations; they are not yet among its formal acceptance claims, and we say so. ↩
-
The sanctions-corpus run illustrates the behaviour on real, messy data. We are not publishing its throughput or compression figures until they are packaged as reproducible benchmarks with fixtures and environment records. ↩
-
Nothing here validates any particular model's reasoning. ↩
Continue the technical conversation
Where could this help your work?
Bring a research question, a database workload or an application you want to build. Let’s connect the ideas in this article to an evaluation that matters to your team.
Discuss this work