MizanS1 · Overseas inference leak

ITU AI for Good Lab Hackathon · United Arab Emirates · Government services

A policy war-room for public-service AI incidents.

Six role agents get different fragments of a synthetic incident at different times. They argue with citations to UAE law and global examples, and only a human can decide. Every step is written to a hash-chained audit log, and this page renders that log.

Run boardS1 · baseline
Audit records, SHA-256 chained
Messages
Gaps
KB docs
Chunks

    Synthetic scenario. Agents play institutional roles, not the positions of any real UAE entity. Agents are deterministic and retrieval-grounded (no LLM in this build). Not legal advice.

    01 · Replay

    War-room replay

    T+70
    Audit chain—
    Records—
    Head hash · SHA-256—

      02 · Gaps

      Gap register

      Domain-level gaps found while mapping the incident to the KB, each with global examples. "Not found in KB" means not found in this corpus, not absent from law. Every row needs expert review.

      03 · Alignment

      Y.3172 pipeline and AI Readiness mapping

      Each message carries the Y.3172 node it acts on and the ITU AI Readiness dimensions it touches. Select a node to filter the feed.

      04 · Sources

      Knowledge base

        Manifest

        Not ingested

          05 · Method

          How this works, and where it does not

          Built

          • KB built with the ITU reference ingestion code (parser, chunker, metadata, ChromaDB KnowledgeBase), article-level chunks for laws.
          • Scenario engine releases events per role with delays. The Service Owner only learns about personal data through the Data Protection role.
          • Typed protocol: INFORM · REQUEST · ASSESS · PROPOSE · OBJECT · DECIDE. Only the human may DECIDE.
          • Citation verifier: SUPPORTED, WEAK, UNCITED or INVALID for every claim.
          • Append-only JSONL audit, SHA-256 chained, re-verified in your browser.

          Limits

          • Agents follow hand-written playbooks. Retrieval picks the cited sources, but the claim wording is a template. No LLM reasoning is shown.
          • "Reasoning trace" means facts, rule, retrieval and choice. It is not a model's hidden chain of thought.
          • The KB is incomplete: NIAF, the National Cloud Security Policy and NCEMA 7000 are not in it.
          • English translations only. The Arabic legal text prevails.
          • The eval queries were written by the author, so the scores are optimistic.