Module 03 / 07 · Sovereign Risk Infrastructure · Narrative Intelligence & Signal Processing
News Analytics Portal.
Financial-news intelligence engine that converts raw narrative flow into structured, audit-graded risk signals. Every article is processed through 12 specialised risk agents — twelve distinct lenses (sentiment, sector, geography, executive action, regulatory, legal, M&A, credit, liquidity, geopolitical, ESG, market-moving event) applied per signal — producing risk label, impact score, confidence, sector, geography, and entity context. Four narrative engines power the decision surface: RiskMirror (live narrative-to-portfolio correlation), Trust Heatmap (longitudinal source-credibility), Headline-to-PnL (signal → market reaction velocity at 5 min / 1 h / 1 d / 5 d), and Multi-Hop Impact Graph (entity dependency cascade). AI inference runs on NVIDIA NIM with NVIDIA NeMo Guardrails as a hard-coded factuality / safety / compliance gate — every LLM verdict passes the gate before it reaches the bus. Data Flywheel turns analyst feedback into prompt refinement, so the system gets sharper with use, not staler. Signals export to the ARIN 22-agent council for cross-domain unified verdict — together with Global Risk Intelligence (Module 02), this module forms the Physical-to-Financial Translation Engine of the Sovereign Risk Infrastructure. Multilingual (EN / RU); distribution to Telegram, Excel, PDF, Email digest, WebSocket.
Status
Pre-client institutional diligence-ready
Module 03 / 07 · Live news.saa-alliance.com · NeMo Guardrails enforced · ARIN council export · Twin-module bridge with Global Risk Intelligence
S2 · News pipeline · capability stats & 5-stage signal flow
Raw narrative in. Audit-graded signal out.
S3 · System overview · narrative intelligence pipeline · news → structured signal → council verdict
Full-spectrum financial-intelligence pipeline. Not a news aggregator.
News & Analytics Portal is the narrative-intelligence and signal-processing engine of the SAA Alliance Sovereign Risk Infrastructure. Unlike conventional news aggregators that provide keyword-filtered feeds, this platform operates as a full-spectrum financial-intelligence pipeline — ingesting raw news from a large multi-source publisher footprint, normalising and deduplicating content, then processing every article through 12 specialised risk agents that extract structured signals: risk labels, impact scores, confidence levels, affected sectors, geographic exposure, executive actions, and entity relationships.
Every signal is enriched by the NVIDIA NIM API with summaries, risk classifications, and impact assessments. NVIDIA NeMo Guardrails enforces factuality, compliance, and safety constraints on every LLM output — this is not a soft policy but a hardcoded refusal layer between the model and the report bus. A dedicated NER stage extracts named entities (companies, countries, sectors, executives, counterparties) for graph construction. LLM responses are cached with configurable TTL to minimise inference cost and persisted with full audit metadata (timestamp, source-version, model identifier) for query and replay.
The Multi-Hop Impact Graph module constructs entity-relationship networks from processed articles, supporting deep traversal for cascade-effect modelling and shortest-path risk propagation, plus a lightweight single-level mode that delivers fast centrality metrics (PageRank, betweenness, degree). The graph answers questions a keyword search cannot: «If Company X defaults, which sectors and counterparties are exposed within two hops?» — turning narrative flow into systemic-risk pathway analysis. Specific graph engine and library choices are operational detail.
All signals export to the ARIN 22-agent decision council for unified cross-domain verdict generation. The Data Flywheel closes the learning loop: analyst feedback on signal quality and verification outcomes feeds back into prompt-refinement and source-trust recalibration — the system gets sharper the more it is used. Together with Global Risk Intelligence (Module 02 — physical-financial digital twins and 8 strategic sub-modules across CIP / SCSS / SRO / ASGI / ERF / BIOSEC / ASM / CADAPT), this module forms the full Physical-to-Financial Translation Engine — narrative on one axis, physical reality on the other, deterministic kernel verdict at the centre.
S4 · Core capabilities · ingestion · 12-agent NIM · 4 narrative engines · magic screen · distribution
Eight capability surfaces. One disciplined signal pipeline.
Signal ingestion & normalisation
Multi-source feed ingestion with HTML content parsing. Automated deduplication using content fingerprinting. Date normalisation across timezone variants. Source trust scoring — each feed carries a reliability weight that propagates into per-signal confidence calculations. Configurable ingestion frequency and source whitelisting.
12-agent AI risk analysis
Every article passes through twelve specialised lenses on the NVIDIA NIM pipeline: structured risk label, impact score, confidence level, sector and geographic attribution, executive-action flag, and executive summary. Responses are cached to prevent duplicate API calls. Per-signal regeneration is exposed for compliance re-runs (POST /generate-analysis/{signal_id}) with full audit-trail anchor.
RiskMirror
Real-time narrative monitoring that reflects how news flow correlates with portfolio risk exposure. Maps incoming signals to existing risk factors, highlighting when news-driven risk amplifies or contradicts the quantitative model. Visualises narrative-to-risk propagation through the dashboard WebSocket feed in real time.
Trust Heatmap
Cross-source credibility visualisation that maps which news sources report consistently with verified outcomes. Source reliability is tracked longitudinally — sources whose signals are later confirmed by market movements receive higher trust weights. Heatmap overlays on the dashboard show information-quality zones across sectors and geographies.
Headline-to-PnL engine
Correlation analysis between news headlines and subsequent price / PnL movements. Maps signal timestamps to four reaction windows (5 min, 1 h, 1 d, 5 d) to measure narrative impact velocity and magnitude. Enables retrospective triage: «Which headlines actually moved markets?» — separating noise from signal across the news flow.
Multi-Hop Impact Graph
Entity-relationship network constructed from extracted entities and relations per article. Multi-hop traversal supports deep dependency analysis, cascade-effect modelling, and shortest-path risk propagation. A lightweight single-level mode delivers fast centrality metrics (PageRank, betweenness, degree). Both modes expose API-driven graph queries and dashboard visualisation.
Magic Screen & Data Flywheel
Magic Screen — narrative-visualisation layer that turns complex signal flows into intuitive visual patterns: real-time signal clustering, topic-drift detection, event-cascade visualisation. Data Flywheel — continuous learning loop where analyst feedback on signal quality feeds back into prompt engineering and source-trust recalibration. Large-scale historical signal analytics via dedicated warehouse.
Export & notifications
Telegram — automated digest delivery and per-signal forwarding to channels and direct messages. Excel — structured signal export with composable filters. PDF — formatted institutional intelligence reports. Email digest — scheduled intelligence summaries. WebSocket — real-time signal push to connected clients for dashboard live updates.
S5 · Narrative discipline & anti-hallucination · NeMo Guardrails · audit trail · data flywheel · twin-module bridge
Four disciplines that make narrative trustable.
What separates institutional narrative intelligence from a glorified news feed: every LLM output is gated, every signal is auditable, every analyst correction sharpens the next prediction, and every verdict is anchored against a deterministic-kernel risk surface.
institutional posture
NVIDIA NeMo Guardrails — factuality · safety · compliance
Every NIM completion passes a hardcoded refusal layer before reaching the signal bus. Factuality (no fabricated entities, dates, numbers), safety (no harmful or biased framings), compliance (no leakage of restricted attributes). Removing this gate is a code-path change visible in audit, not a runtime toggle — the same architectural posture as Governor Always on the kernel side.
Provenance, replay, and compliance reproducibility
Every signal is persisted with timestamp, source citation, NER extraction, full LLM verdict, model identifier, and prompt version. GET /signals/:id returns the complete audit anchor — built for Fed / OCC / FDIC SR 26-2 reproducibility, EU AI Act Article 12 record-keeping, and analyst case-file workflows. Re-run on demand under preserved provenance.
Data Flywheel — analyst feedback → prompt & trust recalibration
Analyst corrections (mis-classified, false-positive, missed-signal) feed back into prompt refinement and source-trust weight recalibration. The system gets sharper with every audited signal — no model staleness, no drift unaddressed. Continuous learning is a quality discipline, not a marketing slogan: every verified outcome lifts (or lowers) the source’s longitudinal trust score on the Trust Heatmap.
Narrative + Physical = Physical-to-Financial Translation Engine
Module 03 (narrative) and Module 02 (physical-financial digital twins) are designed as a twin pair. Narrative signals enrich the 280-factor Unified Stress Report; the physical-state surface validates which narrative signals are real-world consequential vs. social-media noise. Both modules export to the ARIN 22-agent council — the council resolves the cross-domain verdict, and when sub-second risk math is required, calls the ARIN22 deterministic kernel (US-registered name; class-routed, CRN-anchored at the 99.9 tail, MC-challenger-backed). Kernel correctness validates on commodity ARM CPU at ~99 µs single kernel call, constant from D = 4 synthetic to D = 200 on real portfolios with no degradation; GPU is for batch scale — no GPU lock-in.
S6 · Production stack · 10 capability layers · edge → API → persistence → cache → LLM → NER → guardrails → graph → distribution → ops
Capability layers. Vendor and engine choices operational detail.
news.saa-alliance.com
| Layer | Capability |
|---|---|
| Edge / Web | Hardened reverse proxy with TLS, HTTP/2, and live WebSocket transport |
| API | Python service exposing typed REST endpoints with horizontal worker scale-out |
| Persistence | Embedded relational store for signals; optional dedicated graph engine for deep traversal |
| Cache | In-memory cache for LLM responses with configurable TTL · cost-discipline layer |
| LLM / AI | NVIDIA NIM API for production inference · multi-model routing |
| NER / NLP | Dedicated entity-extraction stage feeding the Multi-Hop Impact Graph |
| Guardrails | NVIDIA NeMo Guardrails — factuality / safety / compliance filter on every LLM output, hard-coded |
| Impact Graph | Multi-hop traversal for deep cascade analysis; single-level mode for fast centrality (PageRank / betweenness / degree) |
| Distribution | Telegram (digests + per-signal forwarding) · Excel · PDF · Email digest · WebSocket |
| Operations | Containerised service orchestration with health checks and structured logging |
Specific framework, vendor, and engine choices are operational detail and not disclosed publicly.
S7 · Integration capabilities · REST · WebSocket · API-key auth · NDA-bound spec
REST + WebSocket. Full integration spec under standard NDA.
REST & WebSocket transport · API-key authentication · full integration spec released under NDA.
Real-time signal stream
Verified news signals are pushed to connected clients within seconds of ingestion via persistent WebSocket transport. Drop-in for trading desks, monitoring dashboards, and downstream alert pipelines — no polling, no missed events. Reconnect-with-replay supported for client outages.
Filtered signal query
Composable REST query — filter by sector, region, minimum impact, minimum confidence, and date range. Returns ranked, paginated signal lists with entity context. Designed for analyst case-file workflows and back-office batch processing.
Signal detail & provenance
Per-signal endpoint exposes the full LLM analysis, extracted entities, source citation, full audit trail, and the corresponding impact-graph anchor. Built for compliance-grade analyst review and Fed / OCC / FDIC SR 26-2 reproducibility (supersedes SR 11-7 implementation guidance).
On-demand AI analysis
Trigger fresh NVIDIA NIM analysis for any signal under NeMo Guardrails enforcement. Useful for re-running with updated prompt configurations, recomputing on schema changes, and preserving an explicit audit trail of every regenerated verdict.
Impact graph access
Direct query into the entity-relationship graph — multi-hop traversal for cascade and counterparty analysis, plus single-level centrality (PageRank / betweenness / degree) for fast ranking. Both modes return structured graph data for downstream visualisation.
Distribution & notifications
Telegram digests + per-signal forwarding, scheduled email digests, bulk export (Excel / PDF), and ad-hoc retention controls. Multilingual formatting (EN / RU). All distribution channels share the same authenticated audit trail as the core API.
Endpoint paths, auth scopes, rate limits, and retention configuration are operational detail and are released under standard integration NDA — not disclosed publicly.