Query

Use this tool FIRST to plan your query strategy before calling any data tool. Reason about which tools to call, in what order, and why. Identify if tools can run in parallel (no dependencies) or must be sequential (one tool needs another's output).

Search and filter compute infrastructure facilities (data centers, semiconductor fabs, packaging plants, HPC centers). Returns facility details including power capacity, GPU count, owner, status, and location.

Get complete details for a specific facility including all relationships, investments, nearby facilities, supply chain connections through owner org, and compute capacity timeline. Use when you need the full picture of a single facility.

Find facilities within a radius of a geographic point. Useful for geographic clustering analysis or finding what infrastructure exists near a specific facility.

Search for organizations (companies, governments, labs) in the AI compute ecosystem. Returns org profile including financials, type, country HQ, and description. Supports fuzzy name matching via aliases.

Get a complete organization profile: financials (revenue, capex, market cap), facility portfolio (count, total power, total GPUs), supply chain summary (top suppliers, top customers), and investment history. For deep dives on a single company.

Get ranked list of organizations by a metric. For competitive landscape analysis and identifying the biggest players.

Get upstream suppliers and/or downstream customers for an organization. Returns supply chain links with product/service, criticality score (1-10), value, sole-source status, and alternative suppliers.

Identify supply chain chokepoints: organizations that are sole-source suppliers in critical categories, have high criticality scores, or serve many customers with no alternatives. Use for "which categories rest on a sole-source supplier?" or "which suppliers score above 8 on criticality?".

Analyze a country's dependency on foreign suppliers in the AI compute supply chain. Returns total links, foreign dependency percentage, sole-source foreign dependencies, and critical foreign suppliers by category.

Get the complete supply chain graph with all edges. Returns supplier/customer org details, product/service, category, criticality, value, and sole-source status. Use sparingly; it returns a large dataset.

Search investments in AI compute infrastructure: CHIPS Act grants, sovereign AI programs, hyperscaler capex, VC/PE rounds, government loans. Returns amounts, dates, programs, and linked facilities/orgs.

Get investment aggregations over time: monthly totals, counts, and breakdowns by type. For trend analysis of compute infrastructure spending patterns.

Get sovereign AI investment records with program names, investment types, status, recorded USD amounts, announcement/commitment/disbursement dates and estimate flags. The response reports the number of returned investment records.

Get BIS Entity List entries and other export control designations. Returns entity names, countries, designation dates, grounds for listing, and Federal Register citations. Rows may have data_quality_flags (provenance and caveat notes, such as a country derived from the entity address rather than named in the notice, the sourcing of a removal date, or — on rows covered only by inference from BIS's 50% Affiliates Rule — that the rule is suspended and that the ownership evidence behind the inference no longer reproduces from current data) and a coverage_note stating the 50%-Affiliates-Rule suspension status. Results are ordered most-recent designation_date first and capped (default 50 rows, max 200 via limit); count reports the full matching total and metadata.note flags truncation. For recent-changes questions pass designated_after rather than reading the whole list.

Cross-reference BIS Entity List designations against organizations in the Scrutica database. Match methods (match_method field): 4-layer name matching (exact name, alias, Jaccard similarity, trigram fuzzy) plus affiliate-graph closure ("closure" — a corporate ownership/affiliate-graph walk from the designated entity, the majority of live matches: 101/172 as of the 2026-07-15 data snapshot) and a small curated override set. Returns matched organizations with facility count, supply chain connections, match_method, and match confidence score.

What CHANGED in the BIS Entity List — the change log behind /export-controls/changes and its RSS feed. PREFER this over get_bis_designations for freshness questions ('what changed this month', 'latest additions', 'recent removals'): designation rows arrive already grouped into per-Federal-Register-notice change events — one event per printed BIS rule (pinpoint-page citations of one rule merge via the federalregister.gov document-number map), each with its canonical FR citation and link, the notice title and publication date where resolved, a derived event date labeled with its source column, entity/addition/removal counts, per-country counts, a bounded entity-name sample, and the entities cross-referenced to compute-infrastructure organizations. Removal actions are separate date-keyed events (a screening list maintained from additions alone over-blocks forever). Also returns a last-N-ISO-week activity rollup counting BOTH additions and removals — a zero-count week means the Federal Register carried no BIS action that week, not that the week is missing from the rollup. Results are bounded summaries: at most 48 events per call (default 12, newest first) with truncation declared in metadata; cite totals from the count field and per-event entity_count, never by summing truncated lists. Every underlying designation row is anchored to a Federal Register notice (Tier 1); the event date is derived (modal designation_date, falling back to modal effective_date) and labeled with its source.

Get aggregated compute infrastructure statistics by country: facility count, total power MW, total GPU count, total FLOPS, organization count. For country-level comparisons and rankings.

Get high-level platform statistics: total facilities, total organizations, total investments, total power MW, total GPU count, countries covered. For overview and summary responses.

Get count and capacity breakdown by facility type (hyperscale DC, colocation, logic fab, memory fab, packaging, HPC, etc.). For understanding the composition of global compute infrastructure.

Side-by-side comparison of facilities, organizations, or countries. Returns a structured comparison with all relevant metrics aligned. For direct comparison queries.

Search the hardware catalog for GPU, TPU, ASIC, and interconnect specifications. Returns compute performance, memory, TDP, process node, and export control status. For comparing chip specs or identifying which hardware is export-controlled. Half-precision throughput comes back as TWO distinct fields measuring different things — fp16_tflops (dense tensor-core) and fp16_vector_tflops (dense non-tensor vector), which differ by up to 58x on the same part — so name which one you are quoting and never rank across them. Results are ordered by tensor throughput first, then by vector throughput among parts with no published tensor figure.

Search compute deployments: GPU clusters, training runs, cloud regions. Returns hardware type, GPU count, cluster name, training FLOPS, organization, and linked facility. For understanding who is deploying what compute where.

Get cloud GPU pricing data by provider, hardware type, region, and pricing model. Returns price per GPU-hour for comparison across cloud providers.

Query bilateral semiconductor trade flows (HS 8542 integrated circuits). 397K records from UN Comtrade (25 reporters), CEPII BACI (231 reporters), Taiwan Customs, China GACC, and Japan e-Stat. Covers 2010-2026. Use this to answer: "How much did China import from Japan in 2023?", "Show US chip exports to China over time", "Which countries does Taiwan export the most ICs to?"

Query the PJM interconnection queue (6,093 entries). Shows planned and active power projects in PJM territory (Northern Virginia, Ohio, Maryland, Pennsylvania). Use this to answer: "How much power is queued in Virginia?", "What DC-region projects are active?", "Show PJM queue by state."

Get temporal compute capacity snapshots for facilities. Shows how capacity changed over time (GPU additions, power upgrades, new phases). Use this to answer: "When did xAI Colossus add GPUs?", "How has TSMC fab capacity changed?"

Compare the curated model training-compute estimates with numerical components of regulatory instruments. Pass model_name for a model, regime_id for an instrument and models above its numerical line, or neither for cohort counts. Qualitative rules such as the CAC Interim Measures have no numerical crossing or count. Results retain sources, estimate flags, operation types and legal status; they do not establish designation, applicability or compliance.

Rank facilities by nameplate power capacity within the jurisdictions affected by a regulatory regime. Use this for the structural question "which facilities have the physical compute capacity to support a model that triggers regulatory obligations under regime X" — the model-level question is get_threshold_compliance. Returns facilities with power_capacity_mw, total_gpu_count, owner country, and is_estimated flag. The LLM must reason from power capacity to FLOP plausibility itself, not assert "this facility CAN train a 10²⁵ model".

Get the Compute Visibility Index — a per-country capacity-weighted distribution of facility-operator transparency tiers (Tier 1 public-filing UBO; Tier 2 licensed-database-sourced; Tier 3 documented-but-secondary; Tier 4 linked org named, no ancestors documented; Data_Gap no linked org, not yet investigated). Tier 4 is NOT a claim that the chain is concealed: just under half its capacity sits under an organization that declared to the global LEI registry why it reports no consolidating parent, including declarations that no parent exists. Pass country (ISO2) for one country's breakdown; omit for the global summary plus top-10 countries by capacity. The Index is the source of truth for "Compute Under Control" estimates by country. Backed by mv_compute_visibility_by_country and the visibility_index_snapshots table (snapshot citation handle: scrutica:compute-visibility-index:YYYY-MM-DD).

Per-facility tier breakdown for a given country — every facility with its assigned visibility tier (Tier_1 / Tier_2 / Tier_3 / Tier_4 / Data_Gap) and the tier_reason string explaining the classification. Use after get_compute_visibility_index to drill into the facilities driving a country's tier mix. Sorted by power_capacity_mw descending.

Walk an organization's ownership chain upward toward the documented ultimate beneficial owner (UBO). Returns the ordered hops (max 10) with per-hop provenance: data_source on the connecting edge, confidence (HIGH/MEDIUM/LOW), as_of_date, source_class (public_filing / private_database / other), and the opacity classification on termination (terminus_listed / terminus_government / terminus_self_reference / terminus_max_hops / terminus_cycle / terminus_contested — the last meaning sources name different parents at equal confidence+tier and the walk stops without electing one). Use this for "who ultimately owns X" questions and for tracing PE-fund chains to their LP base. Citation: cite each hop's edge_data_source individually — the chain itself is not a single source.

Search the Scrutica sovereign-execution procurement records (USAspending federal contracts, simpler.grants.gov federal solicitations, EU TED tenders, Korean MSIT awards, manual curations). One row per primary record; the record_type column distinguishes solicitation / award / subaward / modification / cancellation. Returns the per-record classifier breakdown (confidence + category + ai_relevance + reasoning) plus the per-component authority tiers (announced / committed / disbursed / deployed_chip / deployed_facility / deployed_utility) so the LLM can cite the MIN-of-inputs composite tier per the classifier methodology. Use for queries like "DoD AI contracts in Q4 2025", "EU Tier-1 frontier-AI tenders", "FedRAMP awards classified as training_compute", "awards over $10M with confidence above 0.9". The records are refreshed weekly from USAspending, daily from the federal-solicitations and EU TED feeds, and by hand for MSIT; each awardee name is resolved against the organization alias table as its row is written.

Per-country, per-stage rollup of sovereign procurement execution (announced → committed → disbursed → deployed). Backed by a rollup table that a database trigger keeps in step with every write to the procurement records, with mean classifier confidence and the sovereign-wealth-fund execution_stage_note joined back from those records at read time. Returns the disclosure_state classification (with_records / disclosed_zero / structurally_undisclosed / no_ingest_yet) so the LLM can frame Saudi PIF / UAE Mubadala / Singapore Temasek / Norway GPFG style structural-non-disclosure correctly rather than as data gap. Composite authority tier is computed MIN-of-inputs per the classifier methodology (a tier-1 facility-evidence claim does NOT inflate a tier-3 chip-allocation press report into a tier-1 deployed assertion). Use for queries like "Saudi sovereign AI execution by stage", "compare announced vs deployed for top sovereign programs", "Korean MSIT FY2024 disbursement". For per-record drill-down use search_procurement_records.

Simulate the structural reach of a hypothetical BIS Entity List addition (or other designation type). Given a seed organization, computes the affiliate-graph closure (downward through SUBSIDIARY/JV/AFFILIATE edges, max 4 hops), the facilities operated/owned by entities in that closure, country distribution, sovereign-LP exposure breakdown (PIF, Mubadala, Temasek etc. with disclosed dollar amounts), and coordination-gap classifications. Returns an ImpactReport: closure size, affected facility count, total power MW + H100-equivalent GPUs, country rollups, sovereign-LP breakdown. NOT a legal interpretation — see caveats[] in the result. NOT a prediction of BIS behavior — structural reach only. Use this for "if BIS designated X, what would be caught" pre-decision analysis.

Run a supply-chain cascade-propagation simulation given a named perturbation (e.g., "TSMC Fab 18 offline 30 days", "ASML disrupted", "Tokyo earthquake"). Returns the propagated impact graph: affected nodes, weighted edge dynamics, inventory-buffer absorption, substitution paths, and historical-backtest comparison rows where the same shock pattern has resolved before. Wraps src/lib/data/cascade-dynamics.ts (buildEdgeDynamics, getInventoryBufferDefs, getSubstitutionPathDefs, HISTORICAL_BACKTESTS) + src/lib/data/cascade.ts (getCascadeGraphData). Provenance: cascade parameters have is_estimated: true; the three historical-backtest rows are the Texas Freeze / Samsung Austin S2 (February 2021), Ukraine neon gas supply (February 2022) and the CoWoS packaging bottleneck (Q2 2023 onward), each with its own source and authorityTier — see HISTORICAL_BACKTESTS at src/lib/data/cascade-dynamics.ts.

Return the structured scenario data for one of Scrutica's five named scenario models. (1) tokyo-earthquake — src/lib/data/tokyo-earthquake-data.ts (memory and equipment locations, authored seismic and compound scenarios, supplier shares and HHI, dated factory accounts and a DDR3 2Gb Q2 2011 contract-price observation; site lists and radii are chosen inputs, with no production-geography percentages or calculated earthquake losses). (2) iran-threat — src/lib/data/iran-threat-data.ts (missile/drone systems, launch sites, air-defense profiles and reported strikes, including damage to AWS facilities in the UAE and Bahrain). (3) scs-cable — src/lib/data/scs-cable-data.ts (selected submarine-cable inventory, landing stations, incident-attribution status, authored rerouting scenarios and chokepoint zones; unsupported project and route records are withheld; cases specify assumed cable sets without capacity-loss, latency or cloud-region exposure estimates). (4) taiwan-strait — src/lib/data/taiwan-strait-data.ts (TSMC fab facilities, substitution paths, Chinese chip production evidence, Japan–South Korea policy chronology and trade research). (5) abqaiq — src/lib/data/abqaiq-data.ts (Saudi data centres and petroleum facilities, assigned substations and grid zones, dated incident accounts and conditional disruption scenarios; unresolved incident-to-site links are null, with no interception ratios or site-hardening classifications). Each scenario has source URLs at the underlying data-module level; cite the specific exported constant when summarising (e.g., "MEMORY_FABS at tokyo-earthquake-data.ts").

Calculate an H100-equivalent power scenario from supplied nameplate MW, using the shared power estimator and its current parameter registry. Returns PUE-adjusted IT load, GPU power/count, throughput, days to a cumulative compute threshold, and all assumptions including MFU and interconnect efficiency. The facility/org fields are caller-supplied labels; no lookup occurs. Retrieve power with get_facility_details first; provide a facility_id or org_name label. Quarter/year verdicts describe 90/365-day scenario windows; this calculation does not establish legal compliance.

Return $/petaFLOP-day broken down by cloud provider, GPU type, region, and pricing model. Computes the normalization across AWS, Azure, GCP, Oracle, CoreWeave, Lambda from src/lib/data/pricing.ts:getCloudPricing() (the cost-index data layer). On-prem TCO 3yr/5yr surfaces alongside as the build-vs-buy boundary. Source per row: provider pricing-page snapshot date + Epoch ML hardware spec (dense BF16 TFLOP/s; no 2:4 sparsity). Filter via params; omit all to return the global rollup. Distinct from get_compute_pricing (which returns raw $/GPU-hour cells) — this returns the petaFLOP-day-normalized, throughput-aware rollup.

Return the methodology block for a Scrutica page. Wraps src/lib/data/page-descriptions.ts:getPageDescription(route) (canonical per-page metadata: title, audience-block prose, methodology, source list, authority-tier breakdown). Use for "how does Scrutica calculate X?" — returns the actual current methodology text rendered on the page, not a paraphrase. Cite the route + the specific methodology section when summarising.

Return one labeled methodology subsection from the /methodology page. Wraps the methodology page's structured data layer + section anchors (e.g., "validation-6", "flop-estimation-bounds", "cascade-criticality-weighting", "threshold-derivation"). Distinct from get_page_methodology (which returns a single page's full methodology); this drills into a specific named subsection within /methodology.

Return the rendered prose content displayed on a Scrutica page. Wraps src/lib/data/page-descriptions.ts:PAGE_DESCRIPTIONS[route] — the canonical per-page metadata (title, short description, long description, audience-block variants for technical / policy users, OG image slug, category, source list). Use for "what does the [page] currently say about [topic]?" — the model retrieves the rendered prose then summarises against the user's question. Distinct from get_page_methodology (which returns just the methodology block); this returns the displayed body content.

Return the N most recent rows from the corrections log — Scrutica's public log of substantive data corrections (wrong-value fixes, methodology revisions, source-attribution corrections). Wraps src/lib/data/corrections.ts:CORRECTIONS. Each entry: id, correction_date, affected_path (route), prior_state, corrected_state, root_cause, source_url, authority_tier. Use this to answer "what's been corrected recently?" with verified post-correction values, NOT to surface uncorrected drafts.

Full-text search the corrections log. Wraps src/lib/data/corrections.ts:CORRECTIONS. Use for "has Scrutica corrected anything about [entity X]?" — searches across affected_path, prior_state, corrected_state, root_cause prose. Distinct from recent_editorial_corrections (chronological); this is keyword-driven.

Return per-data-source freshness state — when each data source was last refreshed, the freshness-class TTL, and the current freshness color (fresh / aging / stale). Wraps src/lib/freshness.ts (FreshnessClass, FRESHNESS_TTLS, computeFreshnessLevel). Freshness classes: regulatory (4h TTL, e.g., BIS Federal Register feeds), market (24h, e.g., trade flows), structural (90d, e.g., facility ownership), narrative (365d, e.g., historical scenarios). Use for "how fresh is the [layer] data?" or "what's overdue for refresh?".

Return the authority-tier provenance metadata for a facility record: data_source, source_url, is_estimated, authority_tier, data_vintage, confidence_score (where applicable). Wraps src/lib/data/data-quality-flags.ts:getDataQualityFlagsForFacility(facilityId). The Tier 1-4 hierarchy: Tier 1 = primary measurement / regulatory filing, Tier 2 = research database, Tier 3 = press / analyst, Tier 4 = inferred / estimated. The Citations Panel UI consumes this output to render per-claim provenance badges.

Return the layer-by-layer anatomy of the AI compute supply-chain chokepoint structure — the "hourglass" shape (many AI labs and materials at the ends, extreme concentration in fab/packaging/equipment in the middle). Wraps src/lib/data/market-shares.ts (HBM_SHARES, ADVANCED_FOUNDRY_SHARES, OVERALL_FOUNDRY_SHARES, EUV_SHARES) plus the rendered concentration summary at /chokepoint-anatomy. Each layer has: concentration_level (moderate / high / extreme), summary_text, vintage, source. Use for "show me the supply-chain concentration map" or "where is the system most exposed?".

Return concentration metrics (HHI, top-N share, severity, governance leverage) for a given product/service category. Wraps src/lib/data/emerging-chokepoints-analysis.ts:getEmergingChokepointAnalysis() (getTopChokepoints, getChokepointsByTrajectory, getChokepointExportData, WEIGHT_SCHEMES, getWeightSensitivity). Categories include: HBM, advanced foundry, EUV lithography, photomask, photoresist, lead-frame substrate, optical interconnect, etc. The Herfindahl-Hirschman Index quantifies market concentration on a 0-10000 scale; HHI > 2500 = high concentration; HHI > 5000 = near-monopoly. Use for "which product categories are most concentrated?" or "how concentrated is HBM3 supply?".

Return a country profile with facility counts, reported power and GPU totals, organisations, sovereign programme funding, supply-chain relationship counts, export-control access and Entity List designations. Accepts a country name or ISO alpha-2 code.

Return Scrutica's active emerging-chokepoint feed — supply-chain constraints currently elevated to watch / elevated / high / critical severity, with deadline-bound risks (e.g., "CoWoS capacity shortfall through Q4 2026", "EU power grid pre-emption Phase 1 begins 2026-09-01"). Wraps src/lib/data/emerging-chokepoints.ts:EMERGING_CHOKEPOINTS (getChokepointsBySeverity, getChokepointsByCategory, getChokepointsWithDeadline). Each entry: id, severity, category, summary, evidence, deadline (where applicable), affected_orgs, governance_leverage. Use for "what supply-chain risks are active right now?" or "what's elevated to critical?".

Search BIS annual country-licensing data (Bureau of Industry & Security export license approvals/denials by destination country, ECCN code, and approval status). 2,587 rows covering CY2018-2022. Wraps src/lib/data/bis-license-actions.ts:getBisOverlayData + getBisCountryEccnDetail. Each row: country (ISO3), year, eccn_code, approval_count, denial_count, return-without-action count, total_value_usd, friction_pct (denial+RWA / total), friction_tier (low/elevated/anomalous). Use for "what was the friction rate for chip exports to [country] in [year]?" or "which countries had anomalous friction on AI hardware ECCNs?". Note: BIS publishes this data annually; CY2023+ pending.

Return quarterly HBM (high-bandwidth memory) merchant market share by manufacturer. Wraps src/lib/data/hbm-market-share.ts:getHbmMarketShareRows() + getHbmCanonicalShareByCell (the per-quarter / per-product-tier / per-manufacturer share table). Manufacturers: SK Hynix, Samsung, Micron. Product tiers: HBM2E, HBM3, HBM3E. Sources: Astute Group, TrendForce, Yole Développement, manufacturer disclosures. Per-cell is_estimated, authority_tier and vintage. Use for "what's the current HBM3E share split?" or "how have HBM shares moved since 2024?". Where a quarter-and-manufacturer cell carries more than one source the divergence is surfaced rather than averaged away, and the canonical-share function resolves the cell by preferring a measured row over an interpolated one, then the lower authority tier, then the more recent vintage.

Return EU electricity grid capacity, HVDC interconnects, and per-country grid signal — the records behind the sovereign-AI grid pre-emption analysis. Wraps src/lib/data/eu-grid.ts:getEuGridChokepointData (PyPSA-eur substations + transmission lines + HVDC interconnects). Per-country: nameplate generation MW, current draw MW, headroom MW, hyperscale-DC announcement count, announced power MW vs grid headroom, pre-emption-risk flag. 16,564 records ingested 2026-05-09 from PyPSA-eur (TSO-disclosed network model). Sources: ENTSO-E, PyPSA-eur, per-country TSO disclosures. Use for "where is the EU grid being pre-empted by hyperscale buildout?" or "what's [country]'s grid headroom for AI compute?".

Return the recent-developments feed — Scrutica's timeline of substantive changes across the records (tenant changes, ownership transfers, capacity milestones, regulatory actions, incidents). Wraps src/lib/data/recent-developments.ts:getRecentDevelopmentsForFacility extended for global-feed access. Each entry: id, effective_date, title, body, data_source, source_url, confidence (high/medium/low), category (tenant_change / ownership_transfer / capacity_milestone / regulatory / incident / other), affected_facility_id (where applicable). Use for "what's changed in the last 30 days?" or "show me [country]'s most recent compute developments".

Return per-model training evidence — the multi-source per-model dataset Scrutica uses to cross-validate FLOP estimates against lab disclosures. Wraps src/lib/data/training-evidence.ts (TrainingEvidencePageData, TrainingClaim, FacilityPowerEvidence, ChipAvailabilityEvidence, FinancialEvidence, SupplyChainEvidence). Each model: training_claim (what the lab said), facility_power_evidence (where it ran, how much MW available), chip_availability_evidence (what hardware available + BIS export-control status), financial_evidence (capex / opex commitments), supply_chain_evidence (HBM / packaging / wafer availability), signal_status (consistent / insufficient / unknown). Use for "is [model]'s claimed FLOP budget consistent with the available facility / chip / supply-chain evidence?". Distinct from get_threshold_compliance (which returns regime-crossing verdict); this returns the underlying evidence chain.