Evidence
Inspect the retrieval, not just the claim.
Reviewed traces show how a request becomes bounded, ranked, source-linked context. Every published example is fixed to a corpus version and an evaluation record.
Claims and evidence
Four claims, each bounded by its evidence.
This is one evaluated EU AI Act workload, not a universal quality claim.
Every figure comes from one frozen evaluation set. It measures that set, not another corpus, language, or workload.
- Claim
Retrieval finds the required evidence on this evaluation set.
Evidence96.2%1Recall@10
LimitFinding is not ranking.
Ordering is measured separately: nDCG@10 82.0%, MRR@10 78.9%.
- Claim
Results point to the official article or annex anchor.
Evidence98.3%1Source-anchor recall@10
LimitFocus is lower than recall.
Citation precision@10 is 53.4%. It measures focus on required anchors, not legal correctness.
- Claim
Unsupported queries return an explicit abstention.
Evidence100%2Abstention accuracy
LimitSmall denominator.
Covers the 10 unsupported fixtures in this set only. It does not establish abstention on another corpus or workload.
- Claim
Every returned passage stayed inside the enforced scope.
Evidence01Boundary violations
LimitFrozen corpus, private endpoint.
Applies to this frozen v1 corpus. The retrieval endpoint is not publicly available.
One published query
How a question becomes cited evidence.
CorpusMesh returns retrieval context for a customer-controlled model. Each result keeps its citation label, EUR-Lex anchor, corpus version, verification scope, and lineage evidence. The service does not generate the final answer.
Caller-controlled
How do Article 6 and Annex III work together to identify high-risk AI systems?
The question enters as text, not as a client-made vector.
- Knowledge base
- eu-ai-act-reference
- Requested filters
- language en · jurisdiction eu
- Result limit
- top_k 10
- Profile
- gemini-embedding-2-768-search-v2
Server-controlled
The request cannot widen its own permissions.
CorpusMesh pins the corpus version, verified state, jurisdiction, language, and authorization class before retrieval.
- Corpus version
- 2026-08-21.1
- Verification
- verified
- Jurisdiction
- EU
- Language
- EN
- Authorization
- internal-preview
- 01Article 6 Classification rules for high-risk AI systemsEuropean Parliament, Council of the European Union#art_6Observed score 0.8504
- 02Article 6 Classification rules for high-risk AI systemsEuropean Parliament, Council of the European Union#art_6Observed score 0.8401
- 03Annex IIIEuropean Parliament, Council of the European Union#anx_IIIObserved score 0.8382
- 04Annex IIIEuropean Parliament, Council of the European Union#anx_IIIObserved score 0.8370
- 05Article 6 Classification rules for high-risk AI systemsEuropean Parliament, Council of the European Union#art_6Observed score 0.8215
Scores are observed retrieval scores, not confidence percentages. The historical request retrieved ten results; five reviewed excerpts are published.
Required citation anchors
Article 6 ranked 1, Annex III at position 3.
- Source set
- 5ac47834fdf9…d48dfffa
- Chunk manifest
- 2d73f00da7bd…81e74924
- Source verified
- 2026-08-214
Published source trace, reviewed 2026-08-213
Inspect the cited EU AI Act tracePublication register
What is published, and what is held back.
One row per artifact. The internal version stays listed so the boundary of what is public remains visible.
80 EU AI Act queries, measured end to end
Inspect the RAG evaluationReview recall, ranking, citation precision, abstention, configuration selection, and sealed reproducibility evidence.
PublishedEU AI Act: Article 6 and Annex III
Inspect the cited EU AI Act traceFollow one measured request through mandatory scope, five ranked passages, official EUR-Lex citations, and sealed lineage.
PublishedVersion 2026-08-24.3 trace
Active internally. Its trace is not authorized for public publication.
Not published
These reviewed public artifacts remain pinned to v1. Version 2026-08-24.3 is active internally, but its trace is not authorized for public publication and the retrieval endpoint is not publicly available.
What this does not prove
Retrieval evidence is not a legal answer.
- The retrieval endpoint remains restricted to an internal bearer token.
- The trace returns source context and does not generate or validate a legal answer.
- The internal reference corpus is not an available commercial knowledge base.
Where each figure comes from.
- Runtime endpoint reportInspect the RAG evaluation
Recorded 2026-08-21 on the deployed private-beta endpoint, 80 responses.
Deployed commit 57c28e321dbddcad4eb57006d4015a5c1e0e204f
Report SHA-256 052cbcd7621f56e6d37b7addd7c931234af2d0f1422948bc3c29210eb2500c7f
- Sealed fixture set and corpus
80 fixed queries on 1,160 document rows, cutoff 10. Split 50 calibration / 30 validation.
Fixture set SHA-256 c1780012a679b35b8d43e32bd8be4149af391be7cd971a66e54d2f348fa65b73
Chunk manifest SHA-256 2d73f00da7bd0cfbf942c26e96dbe27bbcf9161f428dd7985d6d669581e74924
- Published source traceInspect the cited EU AI Act trace
Reviewed 2026-08-21 by CorpusMesh product and source review. Corpus version 2026-08-21.1, 5 of 10 results shown.
Public trace SHA-256 88821ad972ee9c05ed2d56e6f125baa0000c9f9abc4515f34dd82db3ef61fc17
- Official sourceOpen the regulation on EUR-Lex (opens in a new tab)
EUR-Lex, Regulation (EU) 2024/1689, verified 2026-08-21.
These identifiers tie the claims to one evaluated system state. They support provenance; this page does not provide the complete artifacts needed for an independent replay.
Build with bounded knowledge
Need maintained knowledge for your own application?
Tell us which sources, jurisdictions, and update obligations your product depends on. We will use the request to prioritize coverage, not pretend that a planned corpus already exists.
New to retrieval scoring? See how recall@k, MRR and nDCG are calculated.