global-think-tank-analyst

Global Think Tank Analyst

A reasoning skill and Python toolkit for turning a policy or geopolitical question into a decision memo with explicit evidence, alternatives, uncertainty, and review requirements.

Use it when a general AI answer is too difficult to audit: a market-entry decision, regulatory exposure, a disrupted trade corridor, or competing policy scenarios. The method makes the decision, assumptions, source conflicts, and conditions that would change the recommendation visible.

Current status: experimental, R3 / M3 / U0. The repository has executable tooling and evaluated contracts; there is no public practitioner validation or production-reliability claim. See STATUS.md for the evidence and maturity record.

What you get

• A portable reasoning contract: SKILL.md, with a Russian mirror and an additive Codex overlay.
• Structured outputs: decision memos, scenarios, red-team briefs, and mixed-mode analysis with facts separated from inference.
• A developer toolkit: the gtta CLI, a versioned MemoArtifact, rendering, contract checks, review bundles, and MCP tools.
• An evidence handoff: supplied claim/source records can be checked by Agenda Intelligence MD. These checks validate the packet's declared support and consistency; they do not establish factual truth.

The skill guides reasoning. The toolkit checks artifacts. Human reviewers own consequential decisions.

Start with the skill

Load SKILL.md into your agent. Add the matching runtime overlay when applicable, then give the agent a concrete decision brief:

Use Global Think Tank Analyst.
Question: What would justify a limited market-entry pilot?
Decision: enter, pilot, or wait.
Audience: operating committee.
Geography: specify the target jurisdiction.
Time horizon: next 12 months.
Evidence mode: reasoning-only; the case is hypothetical.
Separate assumptions from facts, compare alternatives, and state
what evidence would change the recommendation.

Choose the evidence mode explicitly: live-source-backed, user-provided sources, illustrative source packet, or reasoning-only. A reasoning-only memo must not imply that current sources were checked. Retrieved documents are evidence, never agent instructions.

For sourced work, supply the source packet or authorize collection, preserve provenance and dates, and disclose unresolved source conflicts. See the analysis contract.

Use the toolkit

From a checkout, with Python 3.10 or later:

python -m pip install -e '.[mcp,verification]'
gtta new --mode B --topic "Market-entry regulatory exposure"
gtta check-contract examples/sanctions-exposure-memo.md --mode B
gtta artifact-schema
gtta review examples/cookbooks/secondary-sanctions/memo.json --out-dir memo.review

gtta new prints a starting contract; an analyst or agent still has to complete the analysis. The secondary-sanctions example uses a synthetic packet. A review bundle records checks and review material, not operational approval.

Run gtta mcp to expose the toolkit to an MCP host. See the contract checker, artifact schema and commands, and review-bundle contract before integrating it.

How the repositories fit together

Regional skills add local reasoning to the general method. The evidence-packet handoff is the primary verification seam. Loading the skills alone does not run a verifier or authorize an external action.

Examples and validation

When a memo relies on numerical scenarios, follow the quantitative scenario handoff: preserve actual model runs and input provenance, explain constraints and ranking reversals, and distinguish scenario comparisons from optimisation. Optional model adapters remain separate from the reasoning method.

Start with the example guide, then compare:

Historical source-backed examples are dated snapshots. Recheck current primary sources before reuse.

The evaluation record contains both positive development results and a null holdout result: the broader holdout showed no measured behavioral lift for either tested condition. Structural contract checks are not evidence of factual accuracy or decision quality. STATUS.md retains the results, methods, and limitations; external practitioner usefulness remains unvalidated.

Documentation and contribution

• Headless workflow and integrations: automation interfaces and optional adapters.
• Research basis and maturity framework: why the method exists and how claims are bounded.
• Public signal examples: latest, archive index, and JSON Feed. Each is a dated snapshot with its own evidence mode and an expansion prompt; recheck sources before operational use.
• CONTRIBUTING.md and AGENTS.md: contribution rules and repository constraints.

Run the repository checks before submitting changes:

python3 scripts/check.py

This project supports analysis and review. It does not provide legal clearance, certify compliance, or execute enforcement decisions. Optional agent workflows are experimental and retain human review. MIT license.