Congress background document A/CONF.241/11 (historical dates; current dates use overview)
subject to the principle of human review
PDF · 2
Knowledge from its sources
Evidence and methodology
How sources, interpretations and translations are reviewed, and how errors are corrected before and after publication.
01 — Register the source
Select a public document from an official publisher. Preserve the source URL, collected version and content fingerprint. New sources remain unavailable as answer evidence until source review.
02 — Inspect the evidence
Compare the proposed statement with the actual extracted passage and its page locator. An exact quotation establishes provenance; it does not by itself prove that an interpretation is supported. Source facts, interpretations and proposed actions must remain distinct.
03 — Review both languages
A named reviewer examines the exact Arabic and English revision, its citations and attachments. Arabic paraphrases require their own meaning check. Internal source or report approval does not authorize public release.
04 — Publish the reviewed revision
The owner publishes the exact reviewed content. Editing creates a new draft while the previous published snapshot remains unchanged. Evidence invalidation or withdrawal of reuse rights suspends affected public material.
Knowledge areas
Dalil covers Congress documents, AI laws, policies and strategies in the UAE and GCC, and international frameworks. Coverage depends on official sources being accessible, extractable and reviewed; registration alone does not make a source usable in answers. Jurisdiction and document type matter: a strategy or recommendation is not treated as binding law.
News workflow
Discover news → Prepare a summary → Human review.
Official publishers' links and headlines enter the private workspace. Reviewers may request Arabic and English drafts from a limited excerpt, then verify and edit them. The news summary is separate from its proposed relevance to MOJ. Editorial review does not automatically make news eligible answer evidence; evidence admission and public publication remain separate. Collection, publication and event dates are distinct.
Report workflow
User request → Evidence search → Cited draft → Review and approval.
Dalil searches eligible registered evidence and drafts a cited report. Insufficient evidence is disclosed rather than replaced with an invented answer. Users submit reports for internal review; corrections require a newly reviewed version. Internal approval does not automatically authorise public release.
Knowledge updates
Registered sources are checked periodically and content fingerprints compared for changes. Previous versions and review records are retained; changed versions require review before use. Extraction gaps and access failures remain explicit. News discovery checks listing metadata and does not guarantee detection of every article-body change; reviewers must check the original before approval.
Illustrative evidence check — not an evaluation result
Source: A/CONF.241/11, PDF page 2, paragraph 6. The exact excerpt below refers to human review.
Supported reading: the passage explicitly refers to human review.
Unsupported example: “This passage proves that Dalil is accurate.”
Why that conclusion does not follow: a recommendation about human review is not a measurement of this system. It establishes neither an accuracy score nor institutional endorsement.
This is an editorial teaching example, not a logged model refusal, an automated semantic-verifier result or a human pilot score.
How corrections work
If a quotation, interpretation or translation is wrong, a reviewer returns it with a reason. Corrections create a new revision requiring review. Public withdrawal and evidence suspension prevent continued access to affected publications; historical records remain private.
Source freshness and publication approval are separate checks. A successful download or matching fingerprint alone cannot establish factual completeness or validity.
What has not been established
Citation checks and software tests do not establish professional usefulness, freedom from bias or time savings. Those require separate evaluation. Public pilot results remain “Not yet measured” until reviewed measurements are published. The 95% useful-and-cited objective is a target, not an achieved result.
The published workflow does not claim an independent external audit or UN/UNODC endorsement. Coverage is limited to reviewed sources; some Arabic extraction may be excluded when unreliable.
Data handling and human responsibility
Dalil is hosted on Azure in the UAE. Staff questions and selected permitted public passages may be processed through OpenRouter by an external model provider. UAE hosting does not establish UAE-only inference. Public browsing does not invoke an AI model.
Do not enter sensitive personal, case or unpublished institutional information. Authorized staff remain responsible for deciding how reviewed material is used. Ordinary chat transcripts are not automatically retained as pilot evaluation submissions.