Sentinel · System Architecture

Two channels.
One auditable output.

Stata runs first — estimating five competing models, computing the ECT score, detecting the market regime, and locking all quantitative outputs. Claude receives those fixed values and interprets them. It cannot change a single number. The result is a weekly briefing that is simultaneously rigorous and readable.

The dual-channel architecture

Sentinel is built around a strict separation of functions. The statistical channel is deterministic and reproducible — Stata estimates expanding-window models, scores the current ECT deviation, and assigns a regime before any language model touches the data. Only then does Claude enter, operating strictly as an interpreter of locked outputs — validating, contextualizing, and translating into plain language.

Sentinel dual-channel architecture Headline text flows into two parallel channels — Stata (deterministic, locked first) and Claude (interpretive, reads locked outputs) — which combine into a weekly briefing. 185 headline topic series Daily WORDSTAT scoring · textual data only · no price inputs Channel 1 · Stata Deterministic · locked first 5 competing models (expanding window) ECT z-score · price vs narrative gap Regime: NORMAL / SURPRISE / GEO 20-day cone · Jeffreys bands WoE consensus · decibans outputs locked · Claude cannot alter Channel 2 · Claude Interpretive · reads locked outputs Signal validation · internal consistency Dominant narrative identification Situation awareness · headline context Dörner audit · reasoning error check 3 cone-anchored forward scenarios Weekly briefing Signal dashboard · forecast cone · narrative analysis 93.0% directional accuracy 907 trading days backtest 23.5 Sharpe · M1 champion
93.0%
Directional hit rate · h=1
907
Trading days · expanding-window backtest
23.5
Annualised Sharpe · M1 champion
88%
S&P variance explained by PC-1 narrative
185
Headline topic series · WORDSTAT

Narrative and price share a common trend

The PC-1 narrative composite and the S&P 500 level are formally co-integrated — confirmed by Engle-Granger and Johansen tests. They are the same phenomenon expressed in different units. When they diverge, the ECT z-score measures that gap and the model forecasts mean reversion. The statistical channel locks that signal. Claude interprets what is driving it.

The same architecture — textual data scored and structured, then interpreted by an LLM operating on locked quantitative outputs — scales to any domain where language volume and sentiment carry information about future states. Customer communications, project status reports, regulatory filings, competitive intelligence. Sentinel is the live demonstration.

Available for license

The Sentinel desktop product is available for licensing. The architecture — textual signal processing combined with hybrid cognition interpretation — is also available as a consulting engagement applied to your domain. Every inquiry begins with a complimentary discovery conversation.

Inquire about Sentinel →