Perceptual Macroeconomics · Headline Intelligence · Local RAG

Ask the
headline corpus.
See what moved first.

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 →
768K
Vetted headlines
Aug 2022 – Apr 2026
912
Trading days
in the Sentinel dataset
191
Named topic
frequency series
100%
Local inference —
no cloud API
The charts — Perceptual Macroeconomics visualized

The text moved first.
The price followed.

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.

Narrative vs. S&P 500 — TARIFF Topic · Dual Axis
5-day MA · Aug 2022 – Apr 2026 · Five named regimes
Navy line (left axis): S&P 500 DeviationFromTARGET — daily close minus the Fed-implied 12.94%/yr growth path. Positive = market above target.   Gold line (right axis): TARIFF topic frequency (words per 10,000) in daily headlines, 5-day moving average.   Shaded bands: Five named macroeconomic regimes. The tariff spike at peak 77.6 WP10K coincided with DeviationFromTARGET reaching −544 index points — its lowest value in the sample.
TARIFF Topic Frequency Timeline
Daily WP10K · 5-day MA · Aug 2022 – Apr 2026
TARIFF topic frequency near zero from August 2022 through early 2024 — the topic was present but dormant. Beginning in approximately February 2025, frequency surged sharply to a peak of 77.6 WP10K. This is the narrative precursor to the market inflection documented in Chapter 6 of Human Relevance in an Age of Induction.
Q&A demonstration — grounded analytical answers

Ask a question.
Get grounded evidence.

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.

Sample Q&A — actual system output
"How has tariff and trade coverage shifted in 2025?"
Based on the retrieved headlines, tariff and trade coverage in 2025 shifted dramatically from background noise to a dominant narrative force. From late January 2025 onward, tariff-related headlines surged in frequency, framing trade restrictions primarily as economic risk — with emphasis on retaliatory measures, supply chain exposure, and market impact. The dominant framing was adversarial: tariffs as weapons in a multi-front trade conflict rather than policy instruments.
Retrieved evidence (15 headlines)
2025-02-03
trump hits canada mexico with 25 tariffs china with 10 as trade war begins
2025-03-12
eu prepares retaliatory tariffs on us goods as trade tensions escalate
2025-04-03
markets plunge as trump announces sweeping tariffs on all imports
2025-04-09
trump pauses tariffs for 90 days on most countries except china
2025-05-14
us china reach temporary deal reducing tariffs for 90 days
Actual system output from the local RAG demo. Answer generated by llama3.1:8b from retrieved headline evidence only. No external knowledge or internet access used.
How it works — three machines, one architecture

The Hybrid Cognition
architecture, operational.

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.

Machine 1
Corpus Builder
MATLAB · Offline

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.

Machine 2
Retrieval Engine
FAISS · Local · Deterministic

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.

Machine 3
Interpretation Engine
Ollama llama3.1:8b · Local · Probabilistic

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.

Five named macroeconomic regimes

Regime-aware
narrative analysis.

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.

Fed Tightening
Aug 2022 – Mar 2023
Banking Crisis
Mar – May 2023
AI Boom
Jun 2023 – Dec 2024
Tariff Shock
Jan – Oct 2025
Iran / Oil
Jun 2025 – Apr 2026

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 organizational mirror

The same architecture.
Your data instead.

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.

Sentinel system
Your organization
Daily news headline corpus
Customer support tickets, project status reports, supplier communications
TARIFF, INFLATION, RECESSION FEAR spikes
Risk language, churn signals, attrition indicators, project jeopardy language
DeviationFromTARGET (S&P vs. baseline)
Schedule deviation, churn rate deviation, delivery failure rate vs. baseline
"How has tariff coverage shifted in 2025?"
"What language patterns precede project escalation?" or "What drives customer churn signals?"
See the live demo.
The system runs locally on your machine or in a client demonstration. Contact John Aaron to arrange a live walkthrough — Q&A over the headline corpus, the dual-axis charts, and a discussion of how the architecture applies to your data environment.
Request a demo →