# Case Study: Global Think Tank Analyst

## Key facts
- **Role:** Builder & Maintainer.
- **Current state:** Latest tagged release v1.3.0 (2026-05-09); active `main` includes later documentation and examples.
- **Scope:** A horizontal strategic-risk reasoning contract with six memo modes and four explicit evidence modes.
- **Primary outcome:** Reusable Markdown skills, worked memos, review aids, a public signal archive, and a CI-checked evidence-packet handoff into Agenda Intelligence MD.
- **Evidence:** [GitHub repository](https://github.com/vassiliylakhonin/global-think-tank-analyst), tagged release, examples, and eval materials.
- **Boundary:** No runtime, production-usage, adoption, or benchmark claim; the skill does not retrieve sources or verify facts.

## TL;DR
- Built **Global Think Tank Analyst** — a strategic-risk analysis skill for AI agents that produce policy-risk, sanctions, regulatory, geopolitical, trade, and strategic-risk memos.
- It is a domain reasoning layer (a behavior contract), not an agent framework, validator, MCP server, or eval platform — those concerns live in the companion project Agenda Intelligence MD.
- Forces agents to frame the decision, separate facts from assessments, state uncertainty, reason through actor incentives, identify scenarios, and define watch-next indicators.
- Positioned the project as the **horizontal domain skill** that composes with regional specialists and the Agenda Intelligence MD evidence-packet linter / compatibility runtime.

## Evidence and boundaries
- Public GitHub repository: https://github.com/vassiliylakhonin/global-think-tank-analyst
- Agent-readable orientation file: `llms.txt`.
- Universal agent contract: `AGENTS.md` (project rules: identity, honesty, evidence, naming, definition of done).
- Canonical skill behavior: `SKILL.md`. Codex variant: `codex/SKILL.md`.
- Primary composition contract: [`docs/evidence-packet-handoff.md`](https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/docs/evidence-packet-handoff.md), with a [synthetic packet](https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/examples/evidence-packet-handoff.json) and a standard-library validator wired into CI.
- Worked memos under `examples/` covering sanctions, AML, regulatory, AI policy, trade, energy, CBAM, critical minerals, monetary policy, source-conflict handling, and red-team work — every example labels its evidence mode explicitly.
- Human review aids under `evals/`: review checklist, failure-modes catalogue, starter rubric.
- Public signal archive under `signals/` with a contributor template.

## Project state (self-reported)
- Latest tagged release: **v1.3.0**, 2026-05-09 — composition recipes and live-source-backed examples with Agenda Intelligence MD. Active `main` has continued documentation and example work after the tag.
- Distribution model: plain markdown skill files, attachable to any AI agent. No CLI, no runtime.
- Six memo modes: Quick Brief, Standard Memo, Scenario Brief, Red-Team Challenge, Decision Briefing Pack, and Analyst Training.
- Four canonical evidence modes: `live-source-backed`, `user-provided sources`, `illustrative source packet`, `reasoning-only`.
- Domain coverage: policy risk, sanctions, regulatory, geopolitical, trade, strategic risk.

No production-usage, adoption, or benchmark numbers are claimed.

## Context
- LLMs are good at summarizing geopolitical events. They are weak at turning them into decision-ready intelligence.
- The common failure mode: confident-sounding regional commentary, vague "monitor closely" advice, no decision frame, no actor incentives, no triggers, no evidence boundaries.
- A skill layer needed to be small enough to attach to any agent and strict enough to actually change the output — without becoming a framework or runtime.

## Challenge
Most AI-generated strategic-risk analysis is fluent but decision-light. It rarely says what decision is being supported, separates facts from assessments, states confidence honestly, or names the indicators that would update the judgment.

That is fine for background reading. It is weak for compliance, risk committees, sanctions-exposure decisions, regulatory planning, or any operating decision that has to be defensible.

## Delivery path
- Reframed the project as a **strategic-risk analysis skill for AI agents** — a domain reasoning layer, not infrastructure.
- Wrote `AGENTS.md` as a canonical project-rules spec: identity, honesty rules, evidence rules, naming hierarchy, and definition of done. Agents and contributors read this first.
- Defined the analytical contract every memo must respect:
  - decision framing block (Question / Decision / Audience / Time horizon / Evidence mode);
  - explicit separation of Fact / Assessment / Assumption / Scenario / Unknown;
  - actor incentives and leverage;
  - options with trade-offs;
  - watch-next indicators (concrete, observable);
  - confidence and key unknowns;
  - what evidence would change the judgment.
- Added six **memo modes** matching real-world request shapes: Quick Brief, Standard Memo, Scenario Brief, Red-Team Challenge, Decision Briefing Pack, and Analyst Training.
- Added four **evidence modes** so agents always disclose what their output is grounded in. When live verification is unavailable, the memo must include `EVIDENCE ACCESS LIMITED: no live verification performed in this environment` and lower confidence.
- Wrote four flagship illustrative memos under `examples/` to make the contract concrete.
- Added human review aids under `evals/`: a yes/no checklist, a failure-modes catalogue, and a starter scoring rubric — explicitly labeled as review aids, not a validated benchmark.
- Reframed the public signal archive as **examples of the skill style**, not official intelligence; added `signals/TEMPLATE.md` for contributors.
- Aligned the weekly signal-generation script with the new template and tightened its system prompt against fabrication.
- Repositioned the repo as the horizontal layer: this skill owns the reasoning method, regional specialists own geographic depth, and Agenda Intelligence MD owns the primary evidence-packet check plus compatibility validation surfaces.

## My contribution
- Defined a reusable decision-framing and evidence-mode contract for strategic-risk analysis.
- Separated reasoning behavior from runtime, validation, and source-retrieval concerns.
- Authored memo modes, worked examples, adversarial cases, and human review aids.
- Added explicit stop conditions for legal, sanctions, investment, stale-source, and conflicting-fact scenarios.
- Documented what the skill has not been tested on instead of presenting self-scored examples as validation.

## Outcome
- Frames a broad geopolitical or policy question as a decision problem before producing analysis.
- Forces a clear answer to *what decision this informs*, not just a topic summary.
- Separates facts, assessments, assumptions, scenarios, and unknowns visibly.
- Discloses evidence limits when live verification is unavailable.
- Produces concrete watch-next indicators and decision triggers, not vague "monitor closely" endings.
- Travels across runtimes: ChatGPT, Claude, Gemini, Perplexity, Cursor, Codex, OpenClaw, MCP agents, RAG systems, internal copilots.
- Composes with vertical-specialist skills for region-deep analysis, then exports externally checkable claims and supplied source text to Agenda Intelligence MD for a deterministic packet-completeness check.

## Current update — 2026-07
- **Portfolio role clarified:** this repo is a reasoning-method dependency, not a buyer-facing product surface. Agenda Intelligence MD owns the deterministic claim/source packet check and keeps the older MCP, CLI, A2A, and validation surfaces compatible.
- **Evidence-packet handoff added:** factual and quantitative claims use stable claim IDs and declared source IDs; caller-supplied source text and optional verbatim quotes travel with the packet. Scenarios, assumptions, inferences, and analyst judgments stay in the memo. A synthetic example and CI validator keep the seam runnable.
- **Example surface expanded:** the examples now include multiple live-source-backed memos and source-conflict cases, plus Agenda Intelligence projections for selected examples. They remain examples and review aids, not benchmark or customer evidence.
- **Self-scoring caveat strengthened:** eval docs now explicitly warn that same-family LLM judges can over-mark their own outputs; self-scored deltas are structural sanity checks, not external validation.

## Trust-layer update — 2026-05

Recent additions tightening the skill's behavior on bad inputs. Single-author work; not external validation.

- **Adversarial cases** ([`evals/adversarial/`](https://github.com/vassiliylakhonin/global-think-tank-analyst/tree/main/evals/adversarial)) — three starter stress cases: prompt-injection inside a retrieved source; conflicting load-bearing dates between two sources; op-ed framed as a primary regulatory source. Author-designed traps with per-case pass / fail criteria, not a held-out benchmark.
- **Stop-and-request explicit triggers** ([`SKILL.md`](https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/SKILL.md)) — operational triggers under failure handling: legal / sanctions / investment conclusions; conflicting load-bearing facts; stale time-sensitive sources; personal-level predictions about named individuals; active prompt-injection in retrieved content.
- **Grounding profile in memo template** ([`templates/memo-blank.md`](https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/templates/memo-blank.md)) — final block counting tagged claims by provenance and flagged status. Transparency device, not a score.
- **"What this skill has not been tested on"** ([README Limitations](https://github.com/vassiliylakhonin/global-think-tank-analyst#limitations)) — five honest gaps: no labeled accuracy dataset; no multi-agent trials; no cross-model regression tracking; no live-source automation; limited non-English source coverage.

## What it is not
- Not an autonomous intelligence system.
- Not a factuality verifier or live source retriever.
- Not legal, compliance, sanctions, or investment advice.
- Not a generic agent framework, CLI tool, MCP server, or eval framework.
- Not a benchmarked evaluation framework.
- Not a replacement for human analyst judgment.

## Portfolio context
Global Think Tank Analyst is the **horizontal domain skill** in a portfolio designed to compose:

- **Horizontal domain skill** — Global Think Tank Analyst (this case study). Reasoning method and memo modes, region- and topic-agnostic.
- **Vertical specialist** — [Central Asia + Caspian Hybrid Intelligence Skill](https://github.com/vassiliylakhonin/central-asia-caspian-hybrid-intelligence-skill). Region-deep reasoning for Central Asia, Caspian, sanctions, AML, banking, logistics, energy, and corridor risk.
- **Vertical specialist** — [Gulf + Middle East Hybrid Intelligence Skill](https://github.com/vassiliylakhonin/gulf-middle-east-hybrid-intelligence-skill). Region-deep reasoning for Iran sanctions, GCC finance and energy, sovereign wealth, and maritime chokepoint risk.
- **Evidence-packet checks / compatibility validation** — [Agenda Intelligence MD](https://github.com/vassiliylakhonin/agenda-intelligence-md). Primary claim/source linter plus legacy validation, CLI, MCP, and CI tooling.

This repo does not duplicate its neighbors. Vertical depth lives in vertical-specialist repos. This repo owns the reasoning-to-packet handoff; Agenda Intelligence MD owns the deterministic packet check.

## Why it matters
The skill is small enough to attach to any capable agent, and strict enough to change the shape of the output. The contract does not ask the model to sound smarter; it asks the model to frame the decision, label its evidence, and name what to watch next.

That is the part most generic geopolitical analysis misses.

**decision frame → evidence mode → actor incentives → scenarios → watch-next indicators → human judgment**

## Before / after (illustrative)

> *Excerpt from a live-source-backed example in the repo, condensed for this page. Full memo with sources, scenarios, options, and watch-next indicators: [`examples/live-source-backed-eu-ai-act-simplification.md`](https://github.com/vassiliylakhonin/global-think-tank-analyst/blob/main/examples/live-source-backed-eu-ai-act-simplification.md). Evidence mode: `live-source-backed`.*

**User question:** "What does the May 7, 2026 EU Council–Parliament provisional agreement on AI Act simplification (Omnibus VII) change for our compliance roadmap, and how should we adjust delivery over the next 6 months?"

**Before — generic strategic-risk commentary:**

> The provisional agreement clarifies certain AI Act obligations and indicates a more pragmatic approach to compliance. Companies should monitor the formal adoption process, review their compliance roadmap, and adjust resourcing as needed.

That summarizes the news but does not support a decision. No frame, no evidence boundary, no scenarios, no triggers.

**After — with the Global Think Tank Analyst skill attached:**

- **Decision frame:** Question = roadmap impact; Decision = whether to slow / redirect / accelerate AI Act compliance investment over the next 6 months; Audience = head of legal & compliance and product VP at an AI provider with EU customer exposure; Time horizon = 6 months; Evidence mode = `live-source-backed`; Confidence = Moderate.
- **Key judgment (Moderate):** The agreement **buys schedule, narrows two real obligations, and adds one new prohibition** — but does not relax the underlying AI Act architecture. The dominant move is to **redeploy** compliance budget rather than cut it.
- **Facts (with sources):** Council and Parliament reached a provisional agreement on 2026-05-07 (Omnibus VII); national-sandbox deadline postponed to 2027-08-02; transparency grace period for AI-generated content reduced from 6 to 3 months (new deadline 2026-12-02); new prohibition on non-consensual sexual / intimate content; SMC privileges extended; GPAI supervision clarified with national-authority carve-outs.
- **Assessments (separated from facts):** For providers touching synthetic content, the operative tightening is the 3-month grace period, not the high-risk timeline. The new prohibition will create downstream content-moderation, model-card, and vendor-risk obligations broader than the headline.
- **Scenarios (6 months):** clean adoption (modal); adoption with material amendments; member-state implementation drift; litigation or political shift.
- **Watch-next indicators:** formal adoption notices and OJ publication date; substantive amendments before adoption; Commission delegated/implementing acts on the new prohibition; AI Office guidance on the GPAI carve-outs.
- **What would change the judgment:** material amendment narrowing the synthetic-content relief; a CJEU referral on the new prohibition; Commission guidance interpreting the GPAI carve-outs more broadly than the press release implies.

The skill does not retrieve sources or verify facts — that is the job of a source-backed workflow or [Agenda Intelligence MD](https://github.com/vassiliylakhonin/agenda-intelligence-md). It asks the agent to frame the decision, label its evidence, and name what would change the view.

## Tech stack
- Plain markdown skill files (`SKILL.md`, `codex/SKILL.md`, `AGENTS.md`, `llms.txt`).
- Worked memo examples and human review aids in markdown.
- Lightweight Python helper for the public signal archive.
- GitHub repository with CI validating skill frontmatter.
- Agent-readable `llms.txt` and `signals/index.json` / `signals/feed.json`.

## Relevance
This project demonstrates how I think about useful agent infrastructure: small reusable layers, explicit reasoning contracts, low context cost, honest evidence discipline, and outputs that improve decisions rather than just sounding polished — composed cleanly with vertical specialist skills and a separate infrastructure layer instead of bundling everything into one repo.

## Project links
- GitHub repository: https://github.com/vassiliylakhonin/global-think-tank-analyst
- Latest tagged release: https://github.com/vassiliylakhonin/global-think-tank-analyst/releases/tag/v1.3.0
- Companion infrastructure proof: [Agenda Intelligence MD](/case-study-agenda-intelligence-md.html)
- Vertical specialist proof: [Central Asia + Caspian Hybrid Intelligence Skill](/case-study-central-asia-caspian-skill.html)
- Vertical specialist proof: [Gulf + Middle East Hybrid Intelligence Skill](/case-study-gulf-middle-east-skill.html)