Global Think Tank Analyst (gtta)

An enterprise-grade AI toolkit for strategic-risk analysis, featuring Multi-Agent Debate (MoA) and autonomous Dark Factories.

global-think-tank-analyst turns your AI agents (Claude, ChatGPT, LangChain) into a disciplined geopolitical-risk factory. It enforces evidence separation, uncertainty handling, scenario generation, and outputs structured, decision-ready memos with embedded Knowledge Graphs.

The project has evolved into a complete Python package equipped with an MCP Server, a FastAPI backend, a Streamlit Fleet Control Center, and LangGraph MoA pipelines, making it trivial to deploy advanced analytical reasoning from local laptops to scalable cloud environments.

Read the core analytical prompt (SKILL.md) · See worked examples

No live source retrieval natively. Not legal, compliance, sanctions, financial, or investment advice. Human review is required before operational use.

Quick Start & Installation

Option 1: Docker Compose (Recommended)

The repository includes a ready-to-use Docker configuration that spins up both the FastAPI backend and the Streamlit UI.

git clone https://github.com/vassiliylakhonin/global-think-tank-analyst.git
cd global-think-tank-analyst

export OPENAI_API_KEY="your-api-key"
docker-compose up --build

• UI: http://localhost:8501
• API: http://localhost:8000/docs

Option 2: Python / pip

Install the package via pip:

pip install "global-think-tank-analyst[ui,enterprise,agent]"

Launch the CLI, UI or Server:

# Scaffold a blank memo interactively
gtta new --mode E --topic "EU CBAM Exposure for Kazakh Metals"

# Launch the interactive web UI
export OPENAI_API_KEY="your-api-key"
gtta ui

# Launch the Enterprise REST API
gtta server

Developer Integrations (RAG / AI Agents):
Drop the analyst instructions directly into your agents. We support English and Russian.

from gtta.langchain import get_system_prompt
prompt = get_system_prompt(language="ru", extra_instructions="Focus on logistics.")

Model Context Protocol (MCP) Server:

mcp run src/gtta/mcp_server.py

Try it in one prompt (Zero-Code)

If you just want to use it in ChatGPT or Claude without writing code, attach SKILL.md, then paste:

Use Global Think Tank Analyst.

Question: What does EU CBAM enforcement-phase exposure change for a Kazakh metals exporter over the next 12 months?
Decision this informs: whether to absorb reporting/compliance cost, pass cost through to EU buyers, or restructure export routing.
Audience: founder / operator.
Geography: Kazakhstan and EU import markets.
Time horizon: 12 months.
Evidence mode: reasoning-only unless you can browse live sources.
Depth: quick brief.

Separate facts, assumptions, assessments, scenarios, and unknowns.
Give options, trade-offs, concrete watch-next indicators, confidence, and what would change the judgment.

What it does

• frames broad geopolitical or policy questions as decision problems;
• separates facts, assessments, assumptions, scenarios, and unknowns;
• produces structured strategic-risk memos in seven modes (quick brief, standard, scenario, red-team, decision pack, analyst training, competing hypotheses);
• enforces evidence-boundary language when live verification is not possible;
• helps agents produce concrete watch-next indicators and decision triggers;

What it is not

• not an autonomous intelligence system;
• not a factuality verifier (it structures reasoning, but does not guarantee truth);
• not a live source retriever;
• not legal, compliance, sanctions, or investment advice;
• not a replacement for human analyst judgment.

If you need deterministic checks for source references, declared quotes, lexical support, or unmatched numbers, use the companion project Agenda Intelligence MD and the evidence-packet handoff. Its older memo validation, scoring, and MCP surfaces remain compatibility options.

Portfolio: how this skill composes

Three layers, separately maintained, designed to compose:

For the full portfolio map, see PORTFOLIO.md.

Ecosystem Expansion: You can deploy the Dark Factory engine (API, UI, Agents) from this horizontal skill into any vertical specialist repo using scripts/deploy_engine_to_vertical.py <target_path> <pkg_name>.

flowchart LR
    A[Global Think Tank Analyst<br/>horizontal method] -->|drafts memo| D[Strategic-risk memo]
    V1[V1 Central Asia Caspian<br/>vertical specialist] -->|adds regional depth| D
    V2[V2 Gulf + Middle East<br/>vertical specialist] -->|adds regional depth| D
    D -->|extract factual claims + supplied sources| P[Evidence packet]
    P -->|lint packet completeness| C[Agenda Intelligence MD]
    C --> O[Human review]

Use Global Think Tank Analyst for analyst behavior and memo structure. Bring in a vertical specialist when the region changes the mechanism. Use Agenda Intelligence MD to lint the resulting claim/source packet; treat its result as packet completeness, not factual truth.

This repo does not duplicate either neighbor. Vertical depth lives in vertical-specialist repos; evidence-packet checks live in Agenda Intelligence MD.

This method is loaded as the core reasoning layer in the portfolio's deployed vertical workers. The public browser demos that previously accompanied them are no longer published; examples/ shows the output shape instead.

For the current primary CLI handoff, see docs/evidence-packet-handoff.md and the runnable synthetic examples/evidence-packet-handoff.json. The older memo scoring / MCP recipe remains documented in docs/integrations/agenda-intelligence-md.md as a compatibility workflow; its historical scores are not evidence of current linter performance.

For pre-validation planning, see VALIDATION_PLAN.md, the source-backed public demo docs/case-packet.md, its machine-readable projections (brief, evidence), the docs/headless-workflow.md review path, docs/external-headless-template.md, and the future review-record scaffold in reviews/. These are preparation assets, not evidence of external validation.

Integration status

This repository ships markdown skill files, examples, eval checklists, development-time consistency validators, and a small signal-generation script. It does not include a memo-validation runtime, domain schema layer, or application runtime.

Quick usage

Paste this into any capable AI agent:

Use Global Think Tank Analyst.

Question: [what we need answered]
Decision this informs: [what action depends on it]
Audience: [founder / operator / investor / compliance / policy team / leadership]
Geography: [countries, regions, corridors, markets]
Time horizon: [days / months / 1–3 years]
Evidence mode: live-source-backed / user-provided sources / illustrative source packet / reasoning-only
Depth: quick brief / standard memo / scenario brief / red-team / decision pack / analyst training / competing hypotheses

Separate facts, assumptions, assessments, scenarios, and unknowns.
Give options, trade-offs, indicators to watch, and bounded confidence.

If the agent has live browsing, ask it to cite sources. If it does not, it must say so and lower confidence.

Memo modes

Before / after

A generic LLM answer to "What is the sanctions risk for a European tech firm with operations in Central Asia?":

The geopolitical landscape in Central Asia is complex. Companies should be aware of secondary sanctions risk and monitor developments closely. Engaging stakeholders and remaining agile will be important. Overall the situation requires careful attention.

A Global Think Tank Analyst-style answer to the same question:

Question: Sanctions exposure for a European tech firm operating in Central Asia.
Decision: Whether to expand, hold, or contract Central Asia operations over 12 months.
Evidence mode: reasoning-only.
EVIDENCE ACCESS LIMITED: no live verification performed in this environment.

Key judgment (Moderate confidence): Secondary-sanctions risk is concentrated in dual-use tech and financial-rail exposure, not in general commercial activity. The dominant uncertainty is enforcement posture, not legal text.

Fact: EU and US export-control regimes apply extraterritorially to certain dual-use items.
Assessment: Local banking partners, not the firm, are the primary transmission channel for secondary risk.
Assumption: Current sanctions architecture remains broadly stable for 12 months.
Scenarios: (1) Stable enforcement, narrow exposure. (2) Targeted enforcement against re-export corridors raises compliance cost. (3) Broad financial-rail action forces partner switching.
Actor incentives: Local banks want correspondent access; regulators want optionality; the firm wants predictability.
Options: (a) Maintain with stronger partner due diligence; (b) Ring-fence dual-use SKUs; (c) Restructure payment rails through a compliant hub.
Watch next: New OFAC/EU designations naming regional banks; tightening of re-export thresholds; partner KYC escalation.
What would change the judgment: Direct enforcement against a comparable peer; new dual-use list additions covering core SKUs.

The first answer is a tone. The second is a decision input.

Examples

Latest source-maintenance pass: docs/source-refresh-2026-07-11.md.

Worked memos in examples/, grouped by domain and evidence mode:

For a guided route through the examples, start with examples/README.md.

Live-source-backed examples cite real public sources and each states its own retrieval date (2026-05-08, except the Middle Corridor example at 2026-05-15); verify before any operational use. Reasoning-only examples do not cite live sources and are not intelligence products.

Evaluation

Lightweight, honest review materials in evals/:

• checklist.md — human review checklist
• failure-modes.md — common failure patterns
• rubric.md — starter scoring rubric
• adversarial/ — stress cases: inputs designed to fail predictably (prompt-injection in retrieved sources, conflicting evidence, source mislabeling). The negative counterpart to the checklist.

These are human review aids, not a validated benchmark. For machine-readable validation, scoring, and audit, use Agenda Intelligence MD.

Signal archive

signals/ holds short public examples of the skill style: one signal, why it matters, a bounded assessment, and indicators to watch.

The archive is:

• a public set of examples of how the skill produces output;
• not official intelligence;
• not investment, legal, or compliance advice;
• not a guarantee of factual verification beyond what each signal explicitly cites.

To contribute a signal, copy signals/TEMPLATE.md.

How to consume signals

Use signals as examples of the skill style and as lightweight prompts for deeper memos. They are useful when you want to see how the method handles a current-looking policy, trade, sanctions, regulatory, energy, or macro risk question in compressed form.

Consumption paths:

• read the latest signal at signals/latest.md;
• browse the archive through signals/index.json;
• ingest the JSON Feed at signals/feed.json;
• expand any signal by pasting its "Example expansion prompt" into an agent running this skill.

Signals are not real-time intelligence. Before using one for an operational decision, verify the cited sources and current facts.

Agent-readable endpoints

• AGENTS.md — universal instruction contract for AI agents
• SKILL.md — canonical skill behavior
• codex/SKILL.md — Codex-ready variant
• llms.txt — orientation for LLMs and agent indexers
• signals/index.json — machine-readable signal index
• signals/feed.json — JSON Feed for ingestion

Naming

• Product: Global Think Tank Analyst
• Method: Policy Risk Memo Architect (the analyst behavior the product implements)
• Companion infrastructure: Agenda Intelligence MD

Repository structure

.
├── AGENTS.md             # Universal AI-agent instructions
├── SKILL.md              # Canonical skill behavior
├── codex/SKILL.md        # Codex-ready variant
├── llms.txt              # Agent and LLM indexer orientation
├── docs/                 # Case packets, integrations, review workflow
├── examples/             # Worked memo examples
├── templates/            # Blank memo template for operational use
├── evals/                # Human review checklist, failure modes, rubric
├── reviews/              # Future external review records
├── signals/              # Public signal archive + template
├── scripts/              # Repository checks and signal generation helper
└── .github/              # CI, issue templates, workflows

Roadmap

This project is evolving from a static collection of prompts into a software-driven analytical tool.

Phase 1: Infrastructure & Packaging (Completed)

• Python package gtta and interactive CLI for scaffolding memos.
• MCP (Model Context Protocol) server for seamless IDE/Agent integration.
• Static documentation site generation via MkDocs.
• Localized system prompts (SKILL_RU.md).

Phase 2: Automation & DX (Current)

• Automated syncing: codex/SKILL.md and other format variations will be auto-generated from the root SKILL.md.
• Automated signal indexing: Replacing manual updates to feed.json and index.json with build-time automation.
• CI/CD Evaluations: Integration of promptfoo for LLM-as-a-judge tests on PRs, enforcing evidence discipline.

Phase 3: Advanced AI Integration (Current)

• Framework Adapters: Native SystemMessage classes for LangChain and LlamaIndex.
• Multi-Agent Debate (MoA): gtta.agent now features a LangGraph pipeline with explicit Researcher, Drafter, Critic (Red-Teamer), and Editor nodes to enforce absolute Evidence Discipline.
• Quantitative Execution: The agent pipeline integrates PythonREPLTool for dynamic data fetching and analysis (e.g., pandas/macro API) prior to drafting memos.
• Algorithmic Prompting: Included a dspy-ai pipeline (scripts/optimize_prompt_dspy.py) to systematically optimize the SKILL.md rules against evidence metrics.

Phase 4: Enterprise Capabilities (Current)

• Knowledge Graph Intelligence (GraphRAG): The MoA pipeline extracts entities directly into Mermaid.js knowledge graphs embedded in memos.
• Agentic Memory: gtta.agent supports MemorySaver to preserve context across multiple memo generations (Stateful Sessions).
• Heavy Document Ingestion: The gtta ingest command processes 100+ page PDF reports (e.g. World Bank, RAND) into local FAISS vector stores.
• Production API: A native FastAPI server (gtta server) exposes the multi-agent reasoning pipeline as a scalable REST API.

Phase 5: Autonomous Fleets & Dark Factories (Current)

• Background Agents: Moved from synchronous requests to an asynchronous task queue.
• Fleet Control Center: The gtta ui now features an "Agent Inbox" dashboard. Users can dispatch hundreds of topics simultaneously, spawning parallel background agents.
• Dark Factories: The gtta dark-factory worker runs a continuous autonomous loop. It scrapes breaking geopolitical news, synthesizes risk targets, dispatches the MoA pipeline, and validates output against strict Guardrails (no human review), saving the finalized intelligence directly to signals/autonomous/.

If you'd like to influence the roadmap or contribute to the automation, open an issue.

Contributing

New contributors: CONTRIBUTING.md opens with a "First 15 minutes" onboarding path — read the three load-bearing files (README.md, AGENTS.md, VALIDATION_PLAN.md), run python3 scripts/check.py, and walk one concrete artifact end-to-end. CI runs the same command before changes can merge.

Cross-repo terminology — evidence modes, Axis A/B provenance tags, three-value response logic, and the deliberate maturity-framework asymmetry across the four-repo stack (this repo uses VALIDATION_PLAN.md; vertical specialists use Bar 1/2; agenda-intelligence-md uses ROADMAP.md version targets) — is consolidated in the portfolio glossary at agenda-intelligence-md/docs/glossary.md.

Contact

Author: Vassiliy Lakhonin — Almaty, Kazakhstan (UTC+5).

• Portfolio: vassiliylakhonin.github.io
• Case study for this skill: vassiliylakhonin.github.io/case-study-global-think-tank-analyst.html
• Email: [email protected]
• LinkedIn: linkedin.com/in/vassiliy-lakhonin
• GitHub: github.com/vassiliylakhonin
• Issues and PRs on this repo are welcome.

For external review of an example or the starter rubric (sanctions, regulatory, energy-trading, policy or trade practitioners), please open an issue or email with your background. For bespoke analysis under retainer, email with decision context, geography and time horizon.

License

MIT — see LICENSE.

Paradigm: Dark Factories (Stage 4)

This reasoning engine operates in the Stage 4 paradigm:

• Lingua Franca: Guardrails
• UI: No human review (Headless A2A Engine)
• Agent to Human Ratio: ∞