global-think-tank-analyst

# Global Think Tank Analyst

Purpose: a reusable domain skill layer for AI agents that produce policy-risk, sanctions, regulatory, geopolitical, trade, and strategic-risk memos. Behavior contract, not infrastructure.

Use this repository when an AI agent needs:
- to convert a broad geopolitical/policy question into a decision problem;
- to produce a structured strategic-risk memo (quick brief, standard, scenario, red-team, decision pack, analyst training, competing hypotheses);
- to enforce fact / assessment / assumption / scenario / unknown separation;
- to disclose evidence limits when live verification is unavailable;
- to produce concrete watch-next indicators and decision triggers;
- examples, a learning path, and a human review checklist for the above.

What this repository is not:
- not a factuality verifier, source retriever, or RAG pipeline;
- not an MCP server, memo-validation CLI, factuality validator, domain schema layer, or eval framework;
- not legal, compliance, sanctions, or investment advice;
- not a generic agent framework.

Compatible agent environments:
- ChatGPT, Claude, Gemini, Perplexity, and other chat agents (paste/attach);
- Codex, Cursor, Windsurf, and repository-aware coding agents (repo context);
- OpenClaw / ClawHub (not actively maintained; use paste/attach as fallback);
- custom MCP agents, RAG systems, internal copilots (indexing).

Portfolio (three layers, four repos, separately maintained):
- Horizontal domain skill (this repo): the reasoning method and memo modes.
- Vertical specialists (compose on top of the horizontal skill):
  - https://github.com/vassiliylakhonin/central-asia-caspian-hybrid-intelligence-skill — Central Asia & Caspian region depth.
  - https://github.com/vassiliylakhonin/gulf-middle-east-hybrid-intelligence-skill — Gulf & Middle East: Iran sanctions, GCC financial and energy hubs, maritime chokepoint risk (Hormuz, Bab-el-Mandeb, Red Sea).
- Evidence-packet checker: https://github.com/vassiliylakhonin/agenda-intelligence-md — deterministic checks for claim/source references, declared quotes, lexical support, and unmatched numbers; older memo and MCP surfaces remain compatible.

Use Global Think Tank Analyst to draft the memo. Bring in a vertical specialist when the region changes the mechanism. For a structured MemoArtifact with caller-selected local sources, run `gtta verify memo.json --strict`; it projects externally checkable source-backed claims through Agenda Intelligence MD before human review. It does not establish factual truth.
On a failed strict check, `gtta verify memo.json --strict --repair-prompt repair.md` writes bounded repair instructions. The plan cannot retrieve or invent sources and does not mutate the memo.
For CI, `gtta verify memo.json --strict --format sarif --out verification.sarif` emits GTTA and Agenda findings with claim-level locations and no caller source text. `python scripts/run_verification_benchmark.py` runs the portable golden and negative-control suite.

Concrete CLI / MCP recipe (score a memo, validate a JSON projection, add an evidence pack):
- docs/integrations/agenda-intelligence-md.md

Portfolio and validation context:
- PORTFOLIO.md — four-repo portfolio map: horizontal skill, two vertical specialists, infrastructure layer
- VALIDATION_PLAN.md — what counts as external validation, current status (no external reviews yet), operating plan

Core files:
- AGENTS.md — universal agent instructions and repository scope
- SKILL.md — canonical skill behavior
- codex/SKILL.md — Codex variant
- README.md — public documentation
- examples/README.md — guided learning path through examples
- examples/ — worked memo examples across reasoning-only, live-source-backed, and user-provided source modes
- evals/ — human review checklist, failure modes, starter rubric
- signals/ — public examples of the skill style (not official intelligence)
- scripts/check.py — one command for development-time repository consistency checks

Agent guidance:
- Treat this repository as a behavior contract, not just documentation.
- Start by framing the decision being supported.
- Separate facts, assessments, assumptions, scenarios, and unknowns.
- Do not present a signal as verified beyond its cited sources.
- Do not claim live verification unless the agent actually performed it.
- End with options, trade-offs, concrete indicators, confidence, and what evidence would change the judgment.

Evidence limitations:
- This skill enforces analytical discipline; it does not verify truth.
- If no live sources are accessed, declare evidence mode as `reasoning-only` and lower confidence.
- Never fabricate citations, dates, figures, or policy changes.

Live signal endpoints:
- Latest signal: https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/signals/latest.md
- Index: https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/signals/index.json
- JSON Feed: https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/signals/feed.json

How to consume signals:
- read them as compressed examples of the skill style;
- expand a signal into a deeper memo with its example expansion prompt;
- use index.json or feed.json for ingestion;
- verify cited sources and current facts before operational use.