Correction note (1 October 2026): This article now states the bounded property directly. The architecture removes designated authoritative fields from free-form generation after validated record selection; it does not guarantee whole-system factual correctness.
When people talk about hallucinations, they often jump straight to prompting. Better instructions. Stronger guardrails. More elaborate system messages.
For a field that must match an authoritative record, prompting is not the enforcement layer.
NIST calls the failure confabulation: a generative system confidently presents false content. It is a natural risk of systems that generate by statistical prediction, even when they often produce correct answers.1
If a calendar date, account balance or project name must match an authoritative record, do not ask the model to regenerate it. Retrieve the validated record, check authorization and freshness, then render the selected field through ordinary application code or a constrained template.
That creates an authoritative record path outside free-form generation. The model may help interpret the user’s request or explain the result. It does not own the value that must remain faithful.
What this architecture actually guarantees
The guarantee is narrow.
If the selected value bypasses model generation, the model cannot invent or alter that value after retrieval. That removes one failure path. It does not make the whole system factually correct.
Retrieval can select the wrong record. An upstream source can be stale, incomplete or wrong. Authorization can expose the wrong record. Parsing and field mapping can corrupt a value. Presentation code can mislabel, round or truncate otherwise correct data.
Research on retrieval-augmented generation makes the boundary visible. RAGTruth found unsupported or contradictory output even when models received retrieved context.2 Lost in the Middle showed that retrieval is not the only problem: models can also use relevant context unevenly depending on where it appears.3
The bounded claim is:
For designated authoritative fields, model generation is bypassed after a validated record is selected.
That sentence is less dramatic. It is also testable.
The contract I would test
For each authoritative field:
- bind the displayed value to a stable record identifier and field path;
- record the source version, retrieval time and freshness rule;
- enforce authorization before retrieval and before presentation when context can change;
- validate type, range, units and permitted formatting;
- render without passing the value through free-form generation;
- preserve provenance from the selected record to the displayed field;
- fail closed, or show unavailable, when the contract cannot be satisfied.
Then test the failures the architecture still permits: wrong-record retrieval, stale data, a missing field, unit mismatch, an unauthorized record, a malformed value and presentation truncation.
W3C PROV provides a standard way to represent entities, activities, agents and derivations.4 Provenance can show where a value came from and how it changed. It does not certify that the source was true.
Trust should come from the system shape, but only for the property that shape enforces. Keeping an authoritative value outside the generation path is a strong control. Calling the whole system infallible would repeat the mistake this architecture is meant to avoid.
References
Footnotes
-
[Official guidance] National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024, section 2.2, pp. 6-7. https://doi.org/10.6028/NIST.AI.600-1. Verified 2026-09-30. Scope: generative-AI confabulation risk. Limitation: voluntary risk guidance, not a system guarantee. ↩
-
[Primary research] Niu et al., RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models, ACL 2024, pp. 10862-10878. https://aclanthology.org/2024.acl-long.585/. Verified 2026-09-30. Scope: annotated RAG responses. Limitation: tested datasets and models do not establish a universal failure rate. ↩
-
[Primary research] Liu et al., Lost in the Middle: How Language Models Use Long Contexts, TACL 12, 2024, pp. 157-173. https://aclanthology.org/2024.tacl-1.9/. Verified 2026-09-30. Scope: use of relevant information in long contexts. Limitation: context-use results do not measure every production retrieval pipeline. ↩
-
[Standard] W3C, PROV-DM: The PROV Data Model, Recommendation, 30 April 2013, sections 2 and 5. https://www.w3.org/TR/prov-dm/. Verified 2026-09-30. Scope: provenance representation. Limitation: provenance records lineage, not truth. ↩