Natural-language interface to Scrutica's substrate. The 55 tools the LLM is permitted to call cover facilities, organizations, supply-chain edges, investments, export-control designations, compute deployments, hardware catalogues, pricing, trade flows, the PJM interconnection queue, capacity snapshots, regulatory thresholds, compute-visibility tiers, ownership chains, sovereign-execution procurement, cascade dynamics, scenario models, BIS license actions, HBM market share, EU grid coverage, recent developments, and per-model training evidence. The router prefers Claude Opus 4.7 (1M context); Sonnet 4.6 and GPT-5.5 are the failover. At synthesis time every numerical claim is wrapped in a citation tag, and the panel underneath resolves the tag back to its tool result with an authority-tier badge (Tier 1 primary measurement → Tier 4 inferred), the source URL, and a drill-down on the raw JSON. A post-stream verifier flags any cited record_id the tool-result history cannot account for.
The cross-layer governance questions that today take four datasets and a Python notebook to answer, asked in a chat box: which facilities are operated by BIS-restricted entities, which countries have facilities above the EU AI Act 10²⁵ FLOP-training threshold, how an ASML outage propagates through TSMC's downstream customers, whether the announced sovereign-AI commitments have anything on the ground to show for themselves. Responses are grounded in tool calls against substrate carrying per-claim provenance and an authority tier; the engine refuses rather than fabricates when the answer would require data Scrutica does not hold (no forecasting, no opinions, no real-time prices). Citations export to BibTeX or APA for paste into a paper, and the full conversation exports to markdown.