Global Think Tank Analyst (gtta)

An experimental strategic-risk reasoning framework for AI agents, with versioned memo contracts, CLI, and MCP.

Global Think Tank Analyst turns broad questions about policy, sanctions, regulation, trade, geopolitics, and strategic risk into decision-shaped memos. It makes evidence boundaries, assumptions, uncertainty, actor incentives, options, and watch indicators explicit.

Use the skill · Read the Russian version · See examples · Inspect project status

Current maturity: R3 / M3 / U0. A reproducibly tested package is available from PyPI, alongside an executable method contract and disclosed paired evaluations. No practitioner validation or production reliability is claimed.

GTTA improves analytical structure; it does not establish factual truth. It is not legal, compliance, sanctions, financial, investment, or trading advice. Verify current sources and use qualified human review before acting.

Try it in one prompt

Attach SKILL.md to a capable agent, then paste:

Use Global Think Tank Analyst.

Question: What does regulatory uncertainty change for our market-entry decision?
Decision this informs: enter now, run a limited pilot, or wait.
Audience: operating committee.
Geography: [countries or markets].
Time horizon: 12 months.
Evidence mode: reasoning-only unless live sources are available.
Depth: standard memo.

Separate facts, assumptions, assessments, scenarios, and unknowns.
Tag every material claim with its provenance.
Give options, trade-offs, indicators, confidence, and what would change the judgment.

The skill works without the Python package. This is the simplest and most mature way to use the method.

What it does

• Frames analysis around a concrete decision, audience, geography, and time horizon.
• Separates facts, assessments, assumptions, scenarios, and unknowns.
• Uses per-claim provenance tags: [primary], [secondary], [user-provided], [inference], and [analyst-judgment].
• Calibrates language to evidence and confidence.
• Models actors, incentives, leverage, options, trade-offs, scenarios, and observable triggers.
• Supports seven response modes, from a quick brief to competing hypotheses.
• Exposes the same method through agent instructions, Python, CLI, and MCP.
• Provides deterministic checks for method structure and a strict structured memo artifact for machine-readable workflows.

What it is not

• Not a source-retrieval system or real-time intelligence feed.
• Not a factuality verifier.
• Not an autonomous decision-maker.
• Not a substitute for legal, sanctions, compliance, financial, or domain review.
• Not externally practitioner-validated or production-proven.
• Not a generic multi-agent platform; the optional LangGraph pipeline is an experiment around the core reasoning method.

Install and use

Use the instructions directly

Add AGENTS.md and SKILL.md to an agent workspace, or attach SKILL.md to a conversation. English and Russian instructions are both packaged in the wheel. SKILL.md remains canonical; SKILL_RU.md is a full 45-section Russian rendering with Mode A–G and bilingual canonical output markers so Russian memos remain compatible with the method checker. Structural parity is enforced by scripts/validate_language_parity.py.

Install the developer toolkit from source

The stable package is available from PyPI. A source checkout remains useful for developing the method, examples, and integrations.

git clone https://github.com/vassiliylakhonin/global-think-tank-analyst.git
cd global-think-tank-analyst
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[mcp,verification]"

# Generate a memo scaffold
gtta new --mode B --topic "Market-entry regulatory exposure"

# Heuristically check a Markdown memo
gtta check-contract memo.md --mode B
gtta check-contract memo.md --mode B --format sarif --out gtta.sarif

# Inspect, validate, and render the strict structured artifact
gtta artifact-schema
gtta source-catalog-schema
gtta check-artifact memo.json --json
gtta render-artifact memo.json > memo.md

# Project source-backed claims through Agenda Intelligence and fail closed
gtta verify memo.json --strict
gtta verify memo.json --strict --format html --out review.html
gtta verify memo.json --strict --format sarif --out verification.sarif
gtta verify memo.json --strict --repair-prompt repair.md

# Write the complete strict review contract to a new, auditable directory
gtta review memo.json --out-dir memo.review
gtta check-review-bundle memo.review --json

# Serve the method and artifact tools over MCP stdio
gtta mcp

The current release is v1.8.0:

pip install global-think-tank-analyst

It includes the native Agenda Intelligence verification seam, bounded repair guidance, SARIF output, and a portable verification benchmark while preserving the gtta.memo@1.0 contract. Version 1.8.0 adds gtta review, a single strict orchestration command that atomically writes the versioned review bundle documented in docs/review-bundle.md, plus the Agenda-independent gtta check-review-bundle integrity checker. See STATUS.md for publication state and limitations.

Executable analysis contracts

GTTA separates checks that answer different questions:

Layer	Interface	What it can establish	What it cannot establish
Markdown method preflight	gtta-method-contract@1.x	Required declarations, mode shape, confidence, likely untagged claims, generic advice	Claim boundaries, factuality, source support
Structured memo	gtta.memo@1.0	Claim IDs, provenance, source references, dependency links, mode invariants, canonical rendering	Whether a named source is trustworthy or supports the claim
Memo verification	gtta.memo-verification@1.0 + Agenda Intelligence MD	Native MemoArtifact projection, claim/source packet completeness, declared quotes, lexical support, unmatched numbers	Factual truth or professional approval
Memo repair plan	gtta verify --repair-prompt	Bounded claim-specific repair instructions that preserve unresolved evidence gaps	Source discovery, automatic factual correction, clearance
Review bundle integrity	gtta.review-bundle-check@1.0	Exact file contract, SHA-256 matches, receipt/SARIF shape, internal status consistency	Authorship, authenticity, factual truth, trusted attestation
Operational decision	Human review	Contextual judgment, current-source verification, accountability	Guaranteed correctness

MemoArtifact is the canonical machine-readable GTTA seam. Its claim ledger is shared by the Python API, CLI, and MCP tools; Markdown is the rendered human view.

Verification findings can be emitted as SARIF 2.1.0 with claim-level JSON line locations. The portable six-case benchmark is available through python scripts/run_verification_benchmark.py; CI stores its receipts and uploads regression errors to GitHub Code Scanning. Expected negative-control findings remain in the JSON receipt and do not masquerade as production defects.

Memo modes

Mode	Use it for	Required shape
A — Quick Brief	Fast orientation	Bottom line, risks, watch indicators, confidence
B — Standard Memo	Default decision analysis	Context, actors, assessment, options, change conditions
C — Scenario Brief	Divergent futures	Baseline, scenarios, triggers, implications, indicators
D — Red-Team Challenge	Stress-testing a claim	Target claim, alternatives, failure modes, revised judgment
E — Decision Pack	Team action	Memo, options, watchlist, owner questions, next steps
F — Analyst Training	Developing reasoning	Coaching and Socratic challenge rather than a finished answer
G — Competing Hypotheses	Attribution and rival explanations	Hypotheses, evidence matrix, disconfirmation, sensitivity, bounded judgment

Before and after

A generic answer:

The environment is uncertain. Monitor developments, engage stakeholders, remain agile, and review the strategy regularly.

A GTTA-shaped answer:

Decision: authorize a limited pilot or wait for regulatory clarity. Evidence mode: reasoning-only.

[analyst-judgment] Prefer a reversible pilot because it buys operating information without committing the full rollout budget.

Main downside: delay and duplicated setup cost. Trigger to pause: the regulator expands the authorization requirement to cover the pilot itself. Confidence: Moderate. What would change the judgment: evidence that the pilot creates the same irreversible exposure as a full launch.

The difference is not a more confident tone. It is a visible decision frame, evidence boundary, trade-off, trigger, and revision condition.

How the portfolio composes

GTTA owns the horizontal reasoning method. Regional depth and evidence-packet checks stay in separate repositories.

flowchart LR
    Q[Decision question] --> G[GTTA<br/>reasoning method]
    V[Optional regional specialist] --> G
    G --> M[MemoArtifact / Markdown memo]
    M --> A[Agenda Intelligence MD<br/>evidence-packet checks]
    A --> H[Qualified human review]

Layer	Repository	Responsibility
Horizontal method	Global Think Tank Analyst	Decision framing, memo modes, uncertainty, scenarios, options
Central Asia depth	Central Asia + Caspian skill	Regional mechanisms, corridors, banking, sanctions adjacency
Gulf depth	Gulf + Middle East skill	Gulf banking, energy, maritime chokepoints, Iran-related risk
Evidence packet	Agenda Intelligence MD	Deterministic claim/source packet checks

See PORTFOLIO.md and the evidence-packet handoff for the full seam.

Integration status

Surface	Status	Entry point
Agent instructions	Core	AGENTS.md, SKILL.md, SKILL_RU.md, llms.txt
Python artifact API	Core development interface	gtta.MemoArtifact, check_memo_artifact(), render_memo_artifact(), verify_memo_artifact()
CLI	Tested	gtta new, check-contract, check-artifact, render-artifact, verify
MCP server	Tested optional extra	python -m pip install -e ".[mcp]", then gtta mcp
LangChain / LlamaIndex adapters	Optional	.[langchain] or .[llamaindex]
LangGraph draft-and-critique pipeline	Experimental	.[agent]
FastAPI / Streamlit	Local experiments	.[enterprise,ui]; not a production deployment architecture

Examples

Use examples/README.md as the complete learning path. Start with these:

Goal	Evidence mode	Example
Learn the basic memo shape	reasoning-only	Sanctions exposure memo
See explicit public-source boundaries	live-source-backed	OFAC case memo
See a narrow retrieval boundary	live-source-backed	Middle Corridor logistics risk
Work from supplied documents	user-provided sources	Supply-chain sanctions exposure
Surface conflicting sources	illustrative source packet	IEA–OPEC forecast conflict
Challenge an existing claim	reasoning-only	Red-team policy brief

Every example declares its evidence mode. Source-backed examples are snapshots; verify their retrieval dates and current facts before use.

Evaluation and maturity

The repository contains:

• deterministic regression tests for the CLI, MCP, method checker, and structured artifact;
• an installed-wheel smoke test;
• human review checklists and failure modes under evals/;
• a predeclared 12-case same-task, with/without-skill structural harness under evals/agent-eval/ with an offline Antigravity export/import path and no model API client;
• a versioned declared-behavior extension that keeps structural and behavioral pass rates separate and uses frozen case-specific expectations to move beyond schema-only conformance;
• four published Antigravity runs: two Gemini executions, the seed 20260830 run and a freshness-gated seed 20260831 replication, and two Claude executions, the freshness-gated seed 20260901 cross-model run and fresh post-change seed 20260902 replication; each publishes exact requests, outputs, recorded settings, hashes, mapping, and a deterministic report.

All four completed runs found a 12/12 contract pass rate with the skill and 0/12 for the generic baseline. The Claude run extends the result to a second model family, while its original 181 capped skill warnings expose materially weaker per-claim provenance compliance than the Gemini runs. A narrow gtta-method-contract@1.2.2 precision rescore reduces that stored count to 170 without changing outputs. The fresh post-change Claude replication stores 111 skill warnings, but three samples hit the warning cap, so this is directional evidence rather than a precise causal improvement estimate. These author-operated runs support only a bounded structural-discipline claim: the scorer does not assess factuality, source support, decision quality, or practitioner usefulness. Practitioner review remains U0.

The first preregistered structured declared-behavior run moved beyond the schema-only ceiling: both arms passed 12/12 structural checks, while the skill arm passed 8/12 frozen behavior expectations versus 3/12 for baseline. The observed +41.7 point difference applies only to model-declared artifact fields in one Gemini execution. It is not a factuality, reasoning-quality, causal, or practitioner-usefulness score.

The subsequent Claude Code / Opus 4.6 replication preserved the direction but not the magnitude: 3/12 skill versus 0/12 baseline declared-behavior passes, with 11/12 versus 12/12 structural passes. This is cross-model-family structural evidence, but the low absolute pass rate and one skill invariant failure argue against further headline-score optimization on the same cases.

The preregistered Gemini 3.8 broader-domain holdout then passed strict structure 10/10 in both arms but passed zero complete declared-behavior expectations in either arm. Missing verify: true declarations dominated. This null result does not reproduce the original suite's positive combined-pass difference and is published with its execution qualifications rather than tuned away.

Read STATUS.md for current evidence, docs/maturity-framework.md for the independent release/method/usefulness axes, and docs/definition-of-done.md for claim-specific release gates.

Signal archive

signals/ contains compact examples of the method style. It is not a live intelligence service.

• Latest signal: signals/latest.md
• Machine-readable index: signals/index.json
• JSON Feed: signals/feed.json
• Contribution template: signals/TEMPLATE.md

Re-verify every cited fact before operational use. Any signal can be expanded by running its example prompt through the skill.

Agent-readable endpoints and naming

• AGENTS.md — repository-wide agent contract
• SKILL.md — canonical English runtime instructions
• SKILL_RU.md — full Russian runtime instructions (45 of 45 sections, Mode A–G)
• codex/SKILL.md — Codex-ready variant
• llms.txt — orientation for agents and indexers
• Global Think Tank Analyst — project and horizontal skill
• Policy Risk Memo Architect — analytical method implemented by the skill
• MemoArtifact — versioned machine-readable memo interface

Repository structure

.
├── AGENTS.md                     # Repository contract for agents
├── SKILL.md / SKILL_RU.md        # Canonical runtime instructions
├── STATUS.md                     # Current R/M/U evidence
├── src/gtta/artifact.py          # MemoArtifact schema, validation, rendering
├── src/gtta/discipline.py        # Markdown method-contract preflight
├── src/gtta/cli.py               # CLI adapters
├── src/gtta/mcp_server.py        # MCP adapters
├── docs/                         # Contracts, handoffs, release guidance
├── examples/                     # Worked memos and evidence modes
├── evals/                        # Review material and structural harness
├── signals/                      # Public style examples and feeds
└── tests/                        # Runtime and contract regression tests

Limitations

• GTTA does not retrieve or continuously refresh sources.
• It does not decide whether a source is independent, authoritative, or sufficient for a specific claim.
• check-contract uses Markdown heuristics; use MemoArtifact for exact declared claim accounting.
• check-artifact validates structure and cross-references, not truth.
• The agent pipeline, API, UI, batch jobs, memory, knowledge-graph drafts, and document parsing are experiments, not the release target.
• There is no labeled factual-accuracy benchmark, long-horizon agent trial, or recorded external practitioner review.

Roadmap

1. Keep gtta.memo@1.x and gtta-method-contract@1.x stable, including the machine-readable warning-truncation telemetry added in ruleset 1.2.3.
2. Freeze the completed Gemini and Claude declared-behavior results; do not tune the method or rubric against repeated runs on the same cases.
3. Freeze the completed broader-domain holdout and its null result; do not tune the method or thresholds against those cases.
4. Keep the stable 1.7 verification seam, source catalog, SARIF mappings, and bounded-repair behavior regression-tested.
5. Complete PyPI Trusted Publishing after account access is restored.
6. Record real practitioner feedback if access becomes available; do not use proxy metrics to disguise U0.

Contributing

Read CONTRIBUTING.md, then run:

python3 scripts/check.py

Package changes should also pass the test suite, wheel build, and installed wheel smoke test. Issues and pull requests are welcome.

License

MIT — see LICENSE.