GitHub - vassiliylakhonin/central-asia-caspian-hybrid-intelligence-skill: Regional reasoning specialist for Agenda Intelligence: Central Asia & Caspian skill for AI agents working on corridors, banking, logistics, energy, and geopolitical risk. Specialist layer, not the active commercial wedge.

16 min read Original article ↗

GitHub stars GitHub forks License Validate

CENTRAL ASIA + CASPIAN RISK REASONING SKILL — a reasoning method for AI agents working on regional sanctions, AML, banking, corridor, logistics and energy risk. Open-source. No live data. No legal or compliance advice.

Regional specialist module for Agenda Intelligence. Activated automatically when the analyze tool sees a Central Asia / Caspian geography (Kazakhstan, Uzbekistan, Azerbaijan, Turkmenistan, Kyrgyzstan, Tajikistan, Caspian, Middle Corridor). Also usable standalone via paste/attach into any agent.

1. Vertical specialist: Central Asia + Caspian risk reasoning

This is the vertical specialist skill in the Agenda Intelligence portfolio for Central Asia and the Caspian — a reasoning method that runs inside an AI agent (Claude, ChatGPT, or a custom assistant) and produces mechanism-first, evidence-aware risk analysis on regional sanctions, AML, banking, corridor, logistics and energy questions, instead of the generic regional essays that default LLM output usually returns.

That default weakness is measured, not anecdotal: LLM factual accuracy is systematically lower for regions underrepresented in training data, and generic retrieval agents have been shown to widen that gap rather than close it (arXiv:2503.22877). This is why the skill leans on primary regional sources, currency triggers and per-claim provenance tags instead of the model's prior.

It pairs with a horizontal reasoning skill (Global Think Tank Analyst), a sibling vertical for Iran / GCC / maritime (Gulf + Middle East Hybrid Intelligence Skill), and the Agenda Intelligence MD evidence-packet linter. See section 3 for how they fit together.

It does not replace sanctions screening, AML monitoring, legal review or human analyst judgement. It changes the shape of the reasoning your AI tool produces before any of those steps.

Commercial role

This repo is a regional specialist reasoning layer, not an active commercial wedge. Agenda Intelligence MD is now primarily a deterministic evidence-packet linter; this repo can hand it externally checkable claims and supplied source text without inheriting a buyer claim.

Central Asia / Caspian and Middle Corridor material should support routing, demos, and regional reasoning depth. Do not treat Kazakhstan/local-forwarder or Middle Corridor demand as validated product-market fit. Do not expand buyer-facing positioning, pilot pages, outreach copy, or new workers from this repo unless fresh discovery evidence proves buyer urgency.

2. Who it is for

  • compliance and risk leadership at banks, fintechs and payment providers with SME, correspondent or routing exposure to Kazakhstan, Uzbekistan, Azerbaijan, Turkmenistan, Kyrgyzstan or Tajikistan
  • sanctions desks tracking transshipment, beneficial-ownership opacity and re-export risk across the region
  • logistics, energy and trading firms operating Caspian, trans-Caspian or Middle Corridor routes
  • analysts and researchers covering the region for institutional clients
  • AI builders embedding regional risk reasoning into agents or assistants

3. What you get

  • mechanism-first reasoning: primary driver → transmission channel → exposure
  • explicit uncertainty labels: Verified / Plausible / Judgment / Unknown
  • role-based actions and trigger points — not "monitor closely"
  • an explicit evidence mode and a limitation note on every output
  • no fabricated citations, sanctions designations, or dates

Where this sits in the production AI stack

Reasoning skills (markdown-first reasoning contracts for agents):

  • Global Think Tank Analyst — horizontal: policy, sanctions, regulatory, geopolitical, trade memos
  • → Central Asia + Caspian Hybrid Intelligence Skill (this repo) — vertical: sanctions, AML, banking, corridor risk in Central Asia / Caspian
  • Gulf + Middle East Hybrid Intelligence Skill — vertical: Iran sanctions, GCC banking, sovereign wealth, maritime chokepoint risk

Evidence-packet checker:

  • Agenda Intelligence MD — checks claim/source references, declared quotes, lexical support, and unmatched numbers before human review

The skills define how agents reason. Agenda Intelligence MD reports whether the supplied claim/source packet is complete enough for review. It does not establish factual truth.

Primary handoff: docs/evidence-packet-handoff.md with runnable synthetic examples/evidence-packet-handoff.json.

4. Try it

Paste this into an AI agent using the Codex, Claude or OpenClaw skill file:

Use the Central Asia + Caspian Hybrid Intelligence Skill.

Question: A European fintech is onboarding SME exporters in Kazakhstan and Uzbekistan that trade through Caspian-connected routes. What sanctions, AML, payment-rail and routing risks matter over the next 6-12 months?
Audience: fintech risk and compliance leadership.
Time horizon: 6-12 months.
Evidence mode: reasoning-only unless live source tools are available.
Mode: risk / compliance.

State the primary driver, mechanism, exposure map, role-based actions, trigger points, confidence, unknowns, and limitation note.

Expected shape of a good answer:

  • starts with Primary driver is: ...;
  • explains how the risk transmits through payment rails, counterparties, ownership, routing or corridors;
  • labels uncertainty using Verified / Plausible / Judgment / Unknown where useful;
  • gives role-based actions and trigger points, not vague "monitor closely" advice;
  • includes a clear limitation note and avoids legal or compliance determinations.

Deployed Agenda Intelligence MD profiles built on this regional frame are live as portfolio demos — try them in the browser: Middle Corridor Deal Risk Gate and CIS Secondary-Sanctions Exposure. Evidence triage only, not advice, and not a claim of product-market fit.

5. What it does

This skill helps agents produce mechanism-first, evidence-aware, decision-useful regional risk analysis. It:

  • frames regional questions as concrete risk or strategy problems
  • explains mechanisms before implications
  • separates Verified / Plausible / Judgment / Unknown
  • maps risk transmission channels across banking, payments, routing, ownership and corridors
  • supports sanctions, AML, banking, corridor, energy, logistics and political-economy analysis
  • identifies leverage shifts and actor incentives
  • produces trigger points and watch-next indicators
  • supports role-based implications for banks, fintechs, investors, logistics operators, energy teams and analysts
  • runs a cold-start interview (docs/cold-start-interview.md) to capture role, geography, decision context, risk appetite and source access into a populated practice profile (templates/practice-profile.md) before substantive memos
  • carries an active currency watch (docs/currency-watch.md) listing fast-moving regional topics that source-backed memos should re-verify against current primary sources, with a 90-day staleness rule

6. What it is not

  • not legal advice
  • not compliance advice
  • not sanctions screening
  • not AML transaction monitoring
  • not factuality verification by itself
  • not a live source retriever
  • not a risk database
  • not an agent framework
  • not a CLI, MCP server, or validation platform
  • not a replacement for human analyst or counsel review

7. Relationship to Agenda Intelligence MD and Global Think Tank Analyst

This skill is one of four repos in a wider portfolio. Each has a distinct role; do not blur them.

This repo is the Central Asia / Caspian vertical specialist. Use it standalone or inside the older analyze compatibility workflow. The current primary composition is: horizontal method → regional specialist → claim/source packet → Agenda Intelligence MD linter → human review.

Project maturity. This repo uses a two-bar Definition of Done (Bar 1 — early but credible; Bar 2 — agent-validated specialist resource). Current honest status, per criterion, lives in STATUS.md. Criteria are defined in docs/definition-of-done.md.

  • Central Asia + Caspian Hybrid Intelligence Skill (this repo) — specialist regional/corridor-risk reasoning, domain-specific risk-transmission logic, sanctions / AML / banking / logistics / energy / corridor analysis patterns.
  • Gulf + Middle East Hybrid Intelligence Skill — sibling vertical specialist; reference it when a flow crosses both regions (Iran-Caspian routes, Russia-Iran-China junction, Iraq-Kurdistan corridors, Central Asian energy routed through Gulf hubs). See gulf-middle-east-hybrid-intelligence-skill.
  • Global Think Tank Analyst — broader strategic-risk memo workflow, general policy-risk analysis, scenario and red-team memo modes: https://github.com/vassiliylakhonin/global-think-tank-analyst
  • Agenda Intelligence MD — primary evidence-packet linter; older validation, scoring, CLI / MCP / HTTP / A2A surfaces remain compatible: https://github.com/vassiliylakhonin/agenda-intelligence-md

Use this repo for specialist Central Asia + Caspian reasoning. Use Global Think Tank Analyst for broader strategic-risk memo workflows. Use Agenda Intelligence MD to lint the claim/source packet; treat its result as packet completeness, not factual truth.

This repo does not itself perform Agenda Intelligence MD validation, schema enforcement on outputs, or live source retrieval.

For the full portfolio architecture, see PORTFOLIO.md in Global Think Tank Analyst. docs/companion-patterns.md describes structural patterns for using this skill alongside Agenda Intelligence MD and Global Think Tank Analyst; for current interfaces, schemas and tooling, consult those repos directly.

8. Quick usage

Use the skill variant matching your environment as the operating instruction in your agent setup:

Environment File Notes
Codex runtimes/codex/SKILL.md See also AGENTS.md
Claude runtimes/claude/SKILL.md YAML frontmatter
OpenClaw runtimes/openclaw/SKILL.md
ChatGPT / other LLMs any of the above Paste or attach as system / project instruction

Validation:

python3 scripts/validate.py

The validator checks structure, required phrases, forbidden determinative claims and code-fence balance. It does not validate factuality of any output produced by the skill.

9. Before / after

Before — generic LLM answer:

  • broad regional commentary
  • vague "monitor sanctions"
  • no transmission mechanism
  • no actor incentives
  • no trigger points
  • no evidence boundaries

After — skill-style answer:

  • primary driver
  • mechanism (how the risk transmits)
  • exposure map (where it concentrates)
  • actor incentives and leverage
  • uncertainty labels (Verified / Plausible / Judgment / Unknown)
  • trigger points and watch-next indicators
  • role-based implications (bank, fintech, investor, operator)
  • evidence mode stated explicitly

10. Flagship examples

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

Every example states its evidence mode and ends with a limitation note. The set covers all four canonical evidence modes: six examples use reasoning-only, two use illustrative source packet, six are live-source-backed, and two are user-provided sources.

11. Skill files

  • runtimes/claude/SKILL.md — Claude-compatible variant with YAML frontmatter and Claude-oriented installation wording.
  • runtimes/codex/SKILL.md — Codex-compatible variant with Codex-oriented slug and installation wording.
  • runtimes/openclaw/SKILL.md — OpenClaw-compatible variant with underscore-convention name for ClawHub and install command.
  • docs/cold-start-interview.md — preflight procedure that captures role, geography, decision context, risk appetite, and source access before substantive memo work. STOP rule blocks generic memos when the practice profile is missing or contains [PLACEHOLDER] markers.
  • templates/practice-profile.md — populated profile read by every memo in the session as the default Decision / Audience / Geography / Time horizon block.
  • docs/currency-watch.md — active list of fast-moving regional topics (OFAC Russia/Iran, EU sanctions packages, FATF/EAG status, Middle Corridor, CPC/BTC, etc.) that source-backed memos should re-verify against current primary sources. 90-day staleness rule.
  • scripts/validate.py — dependency-free validator for skill metadata, required sections, source-handling discipline, safety disclaimers, unsafe determinative language and fenced-code balance.

All variants share the same analytical contract: mechanism-first reasoning, evidence labels, role-based implications, trigger points, confidence footer, and explicit limitation notes. Each variant adds platform-specific behavior: Claude variant adds tool-use awareness and evidence-mode shifting; Codex variant adds agentic-loop output discipline and pipeline compression; OpenClaw is the explicit canonical baseline.

Repository layout

.
├── README.md            # Public positioning (this file)
├── AGENTS.md            # Canonical project contract (identity, scope, evidence rules)
├── CLAUDE.md            # Claude Code working rules (inherits AGENTS.md)
├── SKILL.md             # Runtime skill contract
├── STATUS.md            # Honest Bar 1 / Bar 2 status against the Definition of Done
├── CONTEXT.md           # Working context for cross-session continuity
├── CONTRIBUTING.md      # Local validator workflow and CI invariants
├── llms.txt             # Orientation for LLMs and agent indexers
├── runtimes/            # Runtime overlay skill files per platform (claude/, codex/, openclaw/)
├── skills/              # Claude Code plugin packaging (symlink to root SKILL.md)
├── examples/            # Flagship memo examples (state evidence mode)
├── evals/               # Review checklist, failure modes, starter rubric, agent-eval cases
├── docs/                # Source guide, currency watch, cold-start interview, regional logic, risk archetypes
├── templates/           # Practice-profile template populated by the cold-start interview
├── scripts/             # Validator (validate.py)
└── .github/             # CI workflows

Contributing

New contributors: CONTRIBUTING.md opens with a "First 15 minutes" onboarding path — read the three load-bearing files (README.md, AGENTS.md, STATUS.md), run python3 scripts/validate.py locally, and walk one concrete live-source-backed flagship example end-to-end. CI runs the validator on every push; run it locally before opening a PR.

Cross-repo terminology — evidence modes, Verified/Plausible/Judgment/Unknown labels, Axis A/B provenance tags (incl. table-cell discipline), three-value response logic, and the deliberate maturity-framework asymmetry across the four-repo stack (this repo and the Gulf+ME sibling use Bar 1/2; global-think-tank-analyst uses VALIDATION_PLAN.md; agenda-intelligence-md uses ROADMAP.md version targets) — is consolidated in the portfolio glossary at agenda-intelligence-md/docs/glossary.md.

12. Source guide

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

A source-backed workflow uses external retrieval, user-provided source packets, or companion tooling. The skill itself does not retrieve sources.

docs/source-guide.md lists recommended source classes:

  • official sanctions lists (OFAC, EU, UK OFSI, UN, national regimes)
  • national regulators and central banks
  • customs / statistics / trade agencies
  • company registries, BO registers, court records
  • FATF / Egmont / FIU public materials
  • IFIs (IMF, World Bank, EBRD, ADB, AIIB)
  • government releases
  • ports, rail, corridor and logistics operators
  • energy and infrastructure operators
  • credible regional and international media as Tier 2

Listing a source class is not an endorsement and does not guarantee accuracy or timeliness for any specific question.

13. Risk archetypes

docs/risk-archetypes.md catalogues recurring archetypes for the region. For each: mechanism → typical indicators → evidence needed → common false positives → watch-next triggers → role-based mitigation questions. Current archetypes:

  1. Re-export / transshipment exposure
  2. Dual-use goods routing
  3. Correspondent banking de-risking
  4. Beneficial ownership opacity
  5. Sanctioned-party adjacency
  6. Energy / logistics chokepoint risk
  7. Customs / statistics anomaly
  8. State-capacity enforcement gap
  9. China / Russia / EU leverage competition
  10. Caspian corridor disruption
  11. Regulatory arbitrage
  12. Payment-rail exposure

Use them as patterns to structure reasoning, not as factual claims about any specific entity, route or jurisdiction.

docs/regional-logic.md explains when to include which geography. Core rule: do not expand geography for decoration; expand only when it changes the mechanism, risk exposure, leverage or decision.

14. Review checklist

evals/checklist.md — review checklist (not a benchmark) covering scope, reasoning quality, evidence discipline, tone/safety and decision usefulness.

evals/starter-rubric.md — starter rubric for scoring memos against the analytical contract. Honest label: starter rubric, not a validated benchmark.

evals/scoring-example.md — worked scoring examples applying the rubric to a reasoning-only memo and to the live-source-backed memo. The rubric includes evidence-mode-specific dimensions for each of the four canonical modes.

evals/failure-modes.md — common failure modes (generic essay, alarmism without channel, fake sanctions certainty, fake citations, over-expanded geography, missing limitation note, etc.).

Evidence mode vocabulary

Every example and every memo produced with this skill should state one of four canonical evidence modes:

  • live-source-backed — facts retrieved from current authoritative sources at the time of writing.
  • user-provided sources — facts grounded in a source packet supplied by the user.
  • illustrative source packet — facts grounded in a constructed, illustrative source packet for demonstration purposes.
  • reasoning-only — no sources retrieved; structural reasoning only. No factual claims about specific entities, designations or enforcement actions.

15. Limitations

  • This skill helps structure analysis; it does not verify facts on its own.
  • It does not perform sanctions screening, AML transaction monitoring, or live source retrieval.
  • It does not provide legal, regulatory, sanctions, AML, tax, audit or investment advice.
  • It does not guarantee correctness, completeness or timeliness of any output.
  • It is not production-grade risk control infrastructure.
  • Outputs require source-backed verification and qualified professional review before any operational, compliance or commercial use.
  • No production-usage, adoption or benchmark numbers are claimed.
  • No practitioner-validation claim is made.
  • See STATUS.md for an honest status against the Definition of Done in AGENTS.md. The repo currently clears the "early but credible" bar and the "agent-validated specialist resource" bar; it does not claim practitioner validation.

What this skill has not been tested on

Stated honestly so readers can calibrate. These are gaps in observed evidence, not claims of weakness:

  • No labeled accuracy dataset. Adversarial cases in evals/adversarial/ are author-designed traps, not a held-out test set. Pass/fail is judged manually against per-case criteria.
  • No multi-agent or long-horizon trials. Behavior has been exercised in single-turn and short-multi-turn memo production; long autonomous research loops have not been measured.
  • No cross-model regression tracking. Behavior has been observed primarily on Claude. Codex and OpenClaw variants exist but have not been systematically compared head-to-head against the same prompts.
  • No live-source automation. live-source-backed examples were produced with manual source retrieval. There is no integrated retrieval layer here; recency cannot be enforced automatically.
  • Limited non-English source coverage. Russian-, Kazakh-, Uzbek-, and Azerbaijani-language regulatory and registry sources have not been systematically tested as inputs.

16. Roadmap

Indicative direction, not a commitment:

  • continued refinement of docs/risk-archetypes.md indicator catalogues as the field evolves — crypto/VASP archetype (#13) added 2026-05-12
  • additional worked memos in each evidence mode as new use cases arise
  • pairing of live-source-backed memos with primary-list URLs (OFAC SDN, BIS Entity List, EU consolidated, FATF/EAG MER PDFs) once retrieved at point of use

This roadmap is additive. It will not turn this repo into a CLI, MCP server, screening engine or validation platform.

Disclaimer

This repository provides analytical guidance only and does not constitute legal, regulatory, sanctions, AML, tax, audit or investment advice. Outputs require source-backed verification before any operational, compliance or commercial use.