RAG cites sources but still answers wrong, high accuracy still misses cases: how to build auditable test evidence

A knowledge assistant that cites the wrong version, or a classifier at 96% overall that fails in minority languages, is a test-design problem, not a slide-deck problem. ISO 42001 needs data sources, versions, performance by group and refusal mechanisms written up as test evidence and connected to risk treatment and go-live acceptance.

Updated 2026-10-29

Direct answer: a citation does not mean the answer is right, and high accuracy does not mean nothing is missed. A knowledge assistant needs two kinds of failure checked separately: retrieval failure (wrong document or an old version) and generation failure (right data, wrong answer, or a conclusion with no evidence). A classifier needs performance by group, because error rates in minority languages and new customer groups are hidden by the overall average; if test data overlaps training data, the numbers are inflated. What ISO/IEC 42001 requires is that the method, data sources, versions and results of these tests can be traced and connected to risk treatment and go-live approval.

What the test evidence must cover

ProblemWhat must be demonstrable
Citing the wrong versionThe cited document and version can be traced; withdrawn versions are no longer cited
Right data, wrong answerRetrieval and generation failures are counted separately, each with improvement records
Conclusions without evidenceRefusal rules and hand-over to a person exist and have been tested
Accurate overall, misses in a minority groupPerformance by group, cost of errors and go-live thresholds
Test data overlapping training dataData split method and check records
Mixed data sources and licencesA register of source, licensed purpose, personal data, labelling and version; data outside its purpose is re-evidenced, withdrawn or restricted

What a consultant should deliver

How to accept the work

For any release, you can find the test data version, results by group, refusal tests and approval record; for any dataset, you can state its source, licensed purpose and whether it contains personal data.

What cannot be guaranteed

Testing reduces and exposes errors; it cannot guarantee zero hallucination or zero misses. Whether data licensing is lawful is ultimately for legal counsel.

Winners Consulting's approach

The following describes Winners Consulting's own approach. Please test us with the same questions above, and test other consultants too.

Related questions

Request a free mechanism diagnosis

Compliance needs all three; a consultant's value is integration

CRA, TISAX, ISO 27701 and the AI Act are not single technical problems. They are only complete when process, security technology and legal compliance all hold at the same time. Fix only one and the gaps in the other two surface during audits, incident reporting or customer due diligence, and the work has to be redone.

DimensionThe question it answersTypical gap when only this part is done
Process (management system)Who is responsible, what is done, when, where the evidence is, how it improvesTools are in place but nobody runs them on schedule, so there are no records at audit time
Security technologyHow controls are designed and how their effectiveness is verifiedDocuments are complete but the systems do not actually do what they describe
Legal complianceWhich regulation applies, which role you hold, the deadlines and penaltiesControls are built against the wrong regulation or role, so obligations are not covered

Winners' consulting team covers all three dimensions and plans regulatory scoping, process design and technical controls together, so a single body of evidence serves multiple regulations and customer requirements.

Official sources