Skip to content

SRC02-E01 — Three SATs Implemented via LLM

Research R0049 — Landscape Scan
Run 2026-03-31-02
Query Q001
Source SRC02
Evidence E01

Description

Open-source implementation of three structured analytic techniques via LLMs using Streamlit, LangChain, and OpenAI GPT-4.

URL

https://sroberts.io/posts/llm-sats-ftw/

Extract

The implementation covers three SATs: (1) Starbursting — an Idea Generation SAT that develops questions about a topic; (2) Analysis of Competing Hypotheses (ACH) — generates hypotheses and evaluates evidence for/against each, with CSV export for human review; (3) Key Assumptions Check — identifies assumptions in finished intelligence products. Built with Streamlit (Python), OpenAI GPT-4, LangChain, and Pydantic. The author describes it as "basic, but effective" tooling designed for small teams.

Relevance to Hypotheses

Hypothesis Relevance Strength
H1 — Complete prompts exist Contradicts (only 3 of 66 SATs, not a system prompt) Moderate
H2 — No such prompts exist Contradicts (goes beyond narrow tasks) Moderate
H3 — Partial implementations exist Supports Strong

Context

FACT: The implementation is in Python code using LangChain, not as a standalone system prompt. FACT: It covers 3 of the 66 SATs cataloged by Heuer & Pherson. JUDGMENT: This represents the most directly relevant partial implementation for Q001, bridging intelligence analysis methodology with LLM capabilities.

Notes

The human-in-the-loop design (CSV export for analyst review) suggests the author recognized that full automation of analytical rigor is premature. The choice of ACH specifically demonstrates awareness of the intelligence analysis tradition.