ARIN22 application
News Analytics
From headline flow to auditable risk signals.
News Analytics converts narrative flow into structured, auditable support signals for risk teams. AI classifies the narrative, extracts entities and writes summaries; it never computes governed VaR, CVaR or stress numbers.
Impact score is a narrative support signal, not a governed risk metric.
Overview
Not a news aggregator.
The product does not sell a prettier feed. It turns licensed narrative sources into source-cited, replayable support signals that a risk committee can inspect.
Raw text becomes risk context.
Each signal carries source metadata, extracted entities, sector, geography, event class, confidence and impact score as narrative support fields.
AI does not set the risk number.
Narrative signals can enrich ARIN22 and Global Risk Intelligence workflows. Deterministic VaR, CVaR and stress metrics remain owned by the kernel.
Every signal has custody.
Source citation, model identifier, prompt version, extracted entities and analyst review state are preserved for replay and case-file review.
Signal flow
Five steps, no hidden API theatre.
Private routes, auth scopes, rate limits and retention configuration stay in the integration NDA. The public surface explains the workflow, not the private route map.
Ingest
Source-governed publisher footprint, RSS where licensed, timestamp lineage and provenance capture.
Normalize
Deduplication, language handling, entity extraction and event typing before any narrative verdict is formed.
Classify
Twelve specialised lenses produce label, confidence and impact score as narrative support signals.
Gate
NeMo Guardrails checks factuality, safety and compliance before output reaches the report bus.
Export
Signals export to ARIN council and Global Risk Intelligence context with audit metadata attached.
Narrative engines
Four engines explain what changed and why it matters.
The engines are support surfaces. They help analysts see source credibility, market reaction and entity cascades without pretending the LLM is a quantitative model.
Source-to-portfolio context.
RiskMirror maps narrative signals against portfolio, sector and geography context so the team can see what is relevant to their own exposures.
Source credibility over time.
Trust Heatmap tracks source reliability, repetition, corrections and analyst feedback. Trust changes are release-managed, not silent runtime mutation.
Reaction windows, not prophecy.
Headline-to-PnL measures signal-to-market reaction velocity across short windows. It is context for review, not an alpha claim.
Entity cascade and contagion.
Multi-Hop Impact Graph shows entity dependencies, counterparties and sector paths so narrative risk can be routed into the right governed workflow.
Visual clustering without consumer naming.
Signal Canvas replaces the old consumer-facing label. It clusters topic drift, event families and signal bursts for analyst review.
Release-managed learning.
Analyst corrections feed versioned prompt and source-trust calibration releases. If the feedback floor is not met, the release remains pending.
Boundaries
Narrative support is not model validation.
This page is explicit because the module uses generative AI. It is designed for source-cited record-keeping and narrative triage, not for AI-generated risk numbers.
Rights are deployment-specific.
Publisher feeds, RSS sources, redistribution and resale rights are customer and deployment bindings. Public capability statements do not assert universal rights to republish licensed news content or resell derived source material.
No governed risk numbers.
Impact score, confidence and risk label are narrative support signals. VaR, CVaR, stress and portfolio metrics remain deterministic kernel outputs.
Article 12 record-keeping.
The page claims AI-system record-keeping discipline, source citation and replay. It does not claim model-risk validation for a generative or agentic AI module.
No public integration routes.
Private routes, auth scopes, rate limits and retention configuration are operational detail released under standard integration NDA.
Anti-hallucination discipline
The refusal layer is engineered, not promised.
The strongest part of this module is not that it uses AI. It is that the AI output is gated, logged and prevented from masquerading as the deterministic risk engine.
Hard gate before the report bus.
Every completion passes factuality, safety and compliance checks before it becomes a reportable signal. That gate is visible in audit.
Hardcoded refusal layer
NeMo Guardrails is not a soft policy but a hardcoded refusal layer between the model and the report bus.
Code-path visibility
Removing this gate is a code-path change visible in audit, not a runtime toggle hidden in configuration.
Per-signal record
Every signal preserves source citation, NER extraction, model identifier, prompt version and analyst review state.
ARIN council export
Signals can export to the ARIN council, 22 agents, as governed context. The council does not turn AI narrative into a risk number.
Honest scope. Narrative-pipeline disciplines are operational in controlled deployment. Pending first paid pilot: client-specific source-trust calibration, client-side audit ledger integration, per-tenant Data Flywheel state and SOC 2 Type II attestation. The deterministic-kernel side is evidenced on the ARIN22 demo page; this module reuses that evidence rather than re-deriving it.
Delivery
Delivery surfaces with one audit trail.
Output can be delivered into the channels an analyst team already uses, while preserving the same source trail and review state.
Dashboard
Signal list, entity context, trust state and Impact Graph views for analyst triage.
Email digest
Scheduled summaries with source links, risk labels and analyst review status.
Telegram digest
Controlled forwarding for rapid triage, with authenticated audit trail preserved.
Excel / PDF export
Bulk review packs for committees, source-cited case files and internal distribution.
Private integration
Customer integration contract released under NDA. No public integration routes on the site.
EN / RU output
Bilingual formatting where required. Translation is separated from numeric computation and evidence custody.
Pilot entry
Start with the news flow your committee already monitors.
Bring the sources, watchlists, sectors and entities your analysts already follow. SAA returns a source-cited narrative-risk signal pack, claim boundary and replayable audit trail.