The Sentinel Headline Intelligence system queries 768,750 vetted daily news headlines spanning August 2022 through April 2026 — the same corpus that drives the Sentinel forecasting model. Natural-language Q&A grounded in retrieved evidence. Dual-axis charts showing narrative signals co-moving with S&P 500 DeviationFromTARGET. Five named macroeconomic regimes. Fully local — no cloud API.
Request a live demo →The dual-axis chart below plots S&P 500 DeviationFromTARGET (navy, left axis) against TARIFF topic frequency in the headline corpus (gold, right axis). The five named macroeconomic regimes are rendered as colored bands. The tariff spike of early 2025 preceded the price decline by days — the narrative shadow of the market, made visible.
The Q&A system embeds your question into the same 768-dimensional vector space as the headlines, retrieves the most semantically similar evidence from the FAISS index, and generates an answer using only what was retrieved. The model is instructed to say explicitly when evidence is thin or mixed — it never confabulates confidence it doesn't have.
The system enforces a strict separation between the deterministic retrieval layer and the probabilistic interpretation layer. MATLAB builds and vets the corpus. FAISS retrieves evidence by geometric proximity in high-dimensional space. The local LLM interprets only what was retrieved — never the full corpus. This is the Zero Constraint applied: the LLM receives bounded evidence, not open-ended access.
Produces cleaned_daily_headlines_vetted.csv from raw text files. 768,750 vetted headlines across 1,329 trading days. URL-slug cleaning, deduplication, topic categorization. The proprietary knowledge layer no general-purpose AI has access to.
768-dimensional cosine similarity search over nomic-embed-text embeddings. Given the same question, returns the same headlines. Sub-second retrieval. Fully auditable. Implements Kanerva's nearest-neighbor associative memory at scale.
Receives the question plus retrieved headlines only. Generates a grounded analytical answer with explicit evidence citations. Instructed to flag thin or mixed evidence. No cloud API. No data leaves the machine.
The dual-axis chart overlays five named macroeconomic regimes as colored bands across the full date range. Each regime represents a distinct narrative environment — a period when a specific set of topics dominated headlines and drove co-movement with the S&P 500. The regime structure is the same framework used in the Sentinel Perceptual Macroeconomics forecasting model.
The comparative regime analysis feature (planned Phase 3) will allow direct overlay of two selected date ranges — showing how the current narrative environment compares to any prior regime in the corpus.
The 191 named topic series include TARIFF, INFLATION, FEDERAL_RESERVE, RECESSION FEAR, FEAR, DONALD_TRUMP, DOGE, CHINA, RUSSIA & UKRAINE, ARTIFICIAL_INTELLIGENCE, ISRAEL/HAMAS/IRAN, and 180 additional topics spanning politics, markets, geopolitical, economy, and more.
The Sentinel Headline Intelligence system demonstrates the architecture on public data. The same three-machine pipeline applies directly to any organization's internal text corpus — with no changes to the retrieval or interpretation layers.