Back to profile
Case Study: Central Asia + Caspian Hybrid Intelligence Skill
Key facts
- Role
- Builder & Maintainer
- Current state
- Bar 1 and Bar 2 remain cleared under the historical agent-integration canon; as of 2026-07-15 the primary Agenda composition uses a CI-checked evidence-packet handoff
- Scope
- Central Asia and Caspian reasoning across sanctions, AML, banking, corridors, logistics, energy, ownership, and political economy
- Primary outcome
- Canonical skill, runtime overlays, validators, source-freshness rules, adversarial cases, agent-eval records, and a runnable claim/source packet handoff
- Evidence
- GitHub repository,
STATUS.md, validator, and eval directories
- Boundary
- Structural agent validation is not practitioner validation, factual verification, sanctions screening, legal advice, or a benchmark
TL;DR
- Built Central Asia + Caspian Hybrid Intelligence Skill — a vertical specialist skill for AI agents working on sanctions, AML, banking, corridors, logistics, energy and political-economy risk in the Central Asia and Caspian region.
- It is a regional reasoning layer (a behavior contract), not legal advice, sanctions screening, AML transaction monitoring, a CLI, an MCP server, or a validation platform — those concerns live elsewhere in the portfolio.
- Forces agents to reason mechanism-first, label every claim with per-claim provenance tags (source type:
[primary] / [secondary] / [inference] / [analyst-judgment]; action flags: [verify] / [stale-risk: YYYY-MM]), trace risk-transmission channels (banking, payments, routing, ownership, corridors), and produce role-based implications instead of regional essays.
- Positioned as the vertical specialist layer that composes on top of Global Think Tank Analyst, exports externally checkable claims through a small packet, and sends that packet to Agenda Intelligence MD before human review.
Evidence and boundaries
- Public GitHub repository: central-asia-caspian-hybrid-intelligence-skill
- Deployed workers built on this regional frame, with a public browser demo for CIS Secondary-Sanctions Exposure. Evidence triage only, not advice; illustrative usage.
- Universal agent contract:
AGENTS.md (project rules: identity, honesty, evidence, naming, definition of done).
- Skill variants for different runtimes: canonical
SKILL.md plus runtimes/codex/SKILL.md, runtimes/claude/SKILL.md, and runtimes/openclaw/SKILL.md.
- Validator:
scripts/validate.py — checks structure, required phrases, forbidden determinative claims, safety disclaimers, evidence-mode declarations, source-freshness discipline, and code-fence balance. Does not validate factuality of any output.
- Evidence-packet seam:
docs/evidence-packet-handoff.md, a synthetic packet, and a separate standard-library validator wired into CI.
- Status and agent-eval evidence under
STATUS.md, evals/agent-eval/, evals/adversarial/, and evals/skill-improvement/.
- Companion-pattern, source-guide, currency-watch, cold-start interview, and practice-profile documentation under
docs/ and templates/.
Project state (self-reported)
- Distribution: plain markdown skill files, attachable to any AI agent. No CLI, no runtime.
- Implemented layers: canonical skill file, runtime overlays for Codex / Claude / OpenClaw, structural validator, two-bar Definition of Done, source-freshness horizons, cold-start practice profile, adversarial cases, skill-improvement cases, and agent-eval delta records.
- Domain coverage: sanctions exposure, AML, banking and payments, corridor and logistics risk, energy, minerals, ownership tracing, customs/statistics anomalies, beneficial-ownership opacity, and regional political economy across Central Asia and the Caspian.
- Current maturity: Bar 1 and Bar 2 remain cleared under the historical agent-integration canon;
STATUS.md was updated on 2026-07-15 to separate that compatibility evidence from the primary evidence-packet handoff. This is not practitioner validation, factual verification, or a benchmark.
- Current composition: regional memo → selected factual/quantitative claims with stable IDs → declared source IDs and supplied source text → Agenda Intelligence MD packet check → human review. Historical
analyze evals remain compatibility records and do not test the current linter.
- No production-usage, adoption, or benchmark numbers are claimed.
Context
Generic LLMs produce broad, fluent commentary on Central Asia and the Caspian: high-level country narration, vague "monitor sanctions" advice, no transmission mechanism, no actor incentives, no trigger points, no evidence boundaries.
That output is not decision-useful for analysts, banks, fintechs, investors, logistics operators, or energy teams that actually have exposure to the region.
A skill needed to be small enough to attach to any capable agent and strict enough to actually change the shape of regional analysis — without becoming a framework, screening tool, or compliance product.
Challenge
Most AI-generated regional analysis on Central Asia and the Caspian is fluent but decision-light. It rarely traces how a sanction, regulation, or corridor shift transmits into bank exposure, payment routing, ownership chains, or logistics costs. It rarely separates what is verified from what is plausible inference, and it rarely names the trigger points that would update the view.
That is fine for background reading. It is weak for sanctions-exposure decisions, AML escalation, banking risk, corridor planning, energy and minerals exposure, or any regional risk decision that has to be defensible.
Delivery path
- Reframed the project as a vertical specialist regional/corridor-risk skill — explicitly not a generic strategic-memo tool and not infrastructure.
- Wrote
AGENTS.md as a canonical project-rules spec: identity, honesty, evidence, naming hierarchy, and a definition of done with two explicit bars (structural and substantive).
- Defined the regional analytical contract every output must respect:
- mechanism-first reasoning (explain transmission before implications);
- per-claim two-axis provenance tags: source type (
[primary] / [secondary] / [inference] / [analyst-judgment]) plus optional action flags ([verify] / [stale-risk: YYYY-MM]);
- risk-transmission mapping across banking, payments, routing, ownership, and corridors;
- actor incentives and leverage shifts;
- role-based implications (banks, fintechs, investors, logistics operators, energy teams, analysts);
- trigger points and concrete watch-next indicators.
- Authored a canonical
SKILL.md plus Codex, Claude, and OpenClaw runtime overlays from a single analytical contract.
- Added
scripts/validate.py as a structural validator (required phrases, forbidden determinative claims, evidence-mode declarations, source-freshness discipline, safety disclaimers, code-fence balance). Made it explicit that this validator does not check factuality.
- Documented source-freshness horizons so agents disclose how stale a regional reading can be before it stops being useful.
- Added adversarial, agent-eval, and skill-improvement records that list the skill's failure modes and improvement deltas without presenting them as a benchmark.
- Wrote
docs/companion-patterns.md describing how this skill composes with Global Think Tank Analyst (broader memo workflow) and Agenda Intelligence MD (validation, evidence audit, scoring) — illustrative, with a clear pointer to those repos for current interfaces.
My contribution
- Defined a mechanism-first regional reasoning contract for Central Asia and Caspian exposure.
- Built canonical and runtime-specific skill variants from one analytical core.
- Added per-claim provenance tags, freshness horizons, stop conditions, and a structural validator.
- Created adversarial and agent-eval cases that test output structure without presenting them as factual verification.
- Connected the regional layer to the Global Think Tank Analyst method and Agenda Intelligence MD validation surfaces.
Outcome
- Frames a regional question as a concrete risk or strategy problem before producing analysis.
- Forces mechanism-first reasoning across sanctions, AML, banking, corridors, logistics, energy, and political economy.
- Tags every factual claim with per-claim two-axis provenance: source type (
[primary] / [secondary] / [inference] / [analyst-judgment]) plus optional action flags ([verify] / [stale-risk: YYYY-MM]).
- Maps risk-transmission channels and ownership/routing patterns instead of country-level commentary.
- Produces role-based implications (banks, fintechs, investors, logistics operators, energy teams, analysts).
- Names trigger points and watch-next indicators specific to the region.
- Travels across runtimes: Codex, Claude, OpenClaw, and any LLM environment that accepts a markdown skill.
Trust-layer update — 2026-05
Additions tightening behavior under bad regional inputs. Single-author work; does not change STATUS.md — Bar 2 (agent-validated specialist resource) is now cleared per the canon adopted 2026-05-20; the optional practitioner-review trust layer (B2.8) remains open.
- Adversarial cases (
evals/adversarial/) — two starter stress cases drawn from real regional patterns: Caspian STS transfer + re-flagging + AIS-gap framed as "routine logistics"; a Central-Asian bank intermediary marketed as "EU-compliant" for a Russia-routed dual-use trade.
- Stop-and-request explicit triggers (
AGENTS.md) — operational triggers under three-value response logic: definitive legal / sanctions / AML conclusions; conflicting load-bearing facts; stale primary-list references; vessel / cargo claims without verifiable identifiers (IMO, SDN list ID, registry, BO record); active prompt-injection in retrieved content.
- "What this skill has not been tested on" (README Limitations) — five honest gaps including limited testing on Russian-, Kazakh-, Uzbek-, and Azerbaijani-language regulatory and registry sources.
Agent-validation update — 2026-06-30
Bar 2 is now cleared for agent integration, per STATUS.md. This is still not practitioner validation, not factual verification, and not a customer-traction claim.
- Four agent-eval delta cases are recorded under
evals/agent-eval/, including correspondent exposure, customs / mirror-statistics anomaly, Caspian corridor chokepoint, and a false-premise hard-stop case.
- Independent blind judge case: the 2026-06-30 false-premise hard-stop case was scored by independent blind judges (Haiku 4.5 and GPT-5) with the same structural delta. It remains a structural sanity check, not an accuracy benchmark.
- Source-anchored example ratio: 8 of 16 flagship examples are source-anchored, exactly at the Bar 2 threshold. The next added flagship example should be
live-source-backed or user-provided sources to restore margin.
- Cold-start profile: the repo now includes a role / geography / decision-context / risk-appetite interview and practice-profile template so profile-expecting workflows do not default to generic regional memos.
Evidence-packet handoff — 2026-07-15
- Scenarios, assumptions,
[inference], and [analyst-judgment] stay in the memo; externally checkable claims can be exported with stable claim_id and declared source_ids.
- The packet carries supplied source text. A URL, list ID, or citation alone is not treated as source text. Unsupported factual claims remain visible with an empty
source_ids array.
- The synthetic example passes the Agenda Intelligence MD strict CLI and the repo's own CI validator. This establishes interface compatibility, not factual truth or sanctions status.
What it is not
- Not legal advice.
- Not compliance advice.
- Not sanctions screening.
- Not AML transaction monitoring.
- Not a factuality verifier or live source retriever.
- Not a risk database.
- Not an agent framework, CLI tool, MCP server, or validation platform.
- Not a generic strategic-memo tool — that is Global Think Tank Analyst.
- Not a replacement for human analyst or counsel review.
Portfolio context
This skill is the Central Asia + Caspian vertical specialist in a portfolio designed to compose:
- Vertical specialist — Central Asia + Caspian Hybrid Intelligence Skill (this case study). Region-deep reasoning for sanctions, AML, banking, corridors, logistics, energy, minerals, and political economy.
- Horizontal domain skill — Global Think Tank Analyst. Reasoning method and memo modes, region- and topic-agnostic.
- Evidence-packet checks / compatibility validation — Agenda Intelligence MD. Primary claim/source linter plus legacy validation, CLI, MCP, and CI tooling.
This repo does not duplicate its neighbors. The broader memo workflow lives in Global Think Tank Analyst; this repo carries Central Asia / Caspian reasoning depth and the 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 regional output. The contract does not ask the model to sound regionally smart; it asks the model to trace mechanism, label evidence, name the trigger, and say what role the implication is for.
That is the part most generic Central Asia / Caspian commentary misses.
regional question → transmission mechanism → tagged evidence → exposure map → trigger points → role-specific action
Before / after (illustrative)
Excerpt from a flagship example in the repo, condensed for this page. Full memo with full transmission mechanism, exposure map, leverage shifts, and triggers: examples/fintech-sanctions-routing.md. Evidence mode: reasoning-only — operational use requires source-backed verification through official sanctions lists and qualified professional review.
User question: "We are a fintech operating between Kazakhstan and Uzbekistan, with cross-border flows touching Caspian-connected trade routes (Aktau, Baku, Middle Corridor). What are our sanctions, AML, banking and routing risks over the next 6–12 months, and what should we do about them?"
Before — generic regional commentary:
Central Asia faces increased scrutiny over potential sanctions circumvention. Fintechs operating in Kazakhstan and Uzbekistan should ensure robust compliance programs, conduct thorough KYC, and monitor regulatory developments. Working with reputable banking partners and maintaining transparency with regulators is essential.
Fluent regional commentary. Does not say how the risk transmits, who gets affected, or what would update the view.
After — with the Central Asia + Caspian skill attached:
- Bottom line (
[analyst-judgment]): For an SME-facing cross-border fintech in KZ / UZ with Middle Corridor exposure, the dominant risk over 6–12 months is secondary-sanctions and de-risking pressure transmitting through correspondent banking, amplified by routing and ownership opacity in re-export and dual-use flows. The risk is not primarily designation — it is losing or having to re-price USD / EUR rails, and counterparty contamination.
- Risk-transmission mechanism: (1) Western regulators update designations and circumvention guidance → (2) tier-1 correspondents tighten KYC/EDD on KZ/UZ respondents and on flows touching specific HS codes and routes → (3) tier-2 / regional banks pre-emptively de-risk SME exposures → (4) the fintech sees slower settlement, higher rejection rates, narrower BIN/IBAN acceptance, pressure to exit certain corridors.
- Exposure map: banking rails (USD/EUR correspondent access, acquirers, CNY rails); counterparties (SME trading companies with Russia-linked ultimate buyers); routing (Middle Corridor via Aktau–Baku, dual-use HS codes); ownership (opaque BO behind freight forwarders, nominee directors, layered LLPs); geography (Khorgos, Saryagash, UZ logistics hubs, Caspian transshipment).
- Evidence labels visible throughout: sanctions list scope requires direct live screening
[verify]; local AML EDD tightening [secondary][verify]; correspondent de-risking dynamic [analyst-judgment].
- Leverage shifts: tier-1 correspondents and Western regulators gain leverage; SME-facing fintechs without diversified rails and small KZ/UZ banks reliant on a single correspondent lose it.
- Trigger points (watch-next): new OFAC/OFSI/EU designations naming KZ or UZ entities or freight forwarders; public enforcement against a regional bank for circumvention; correspondent-relationship withdrawals affecting tier-2 banks; KZ/UZ regulator updates on BO disclosure or high-risk-HS-code monitoring; customs/statistics divergence on sensitive HS codes; major shift in CNY-rail substitution.
- Role-based implications: different next steps for the fintech itself, its correspondent banks, its SME merchants, and its freight-forwarder counterparties — not one generic "monitor developments."
The skill does not screen sanctions, retrieve sources, or verify facts. It forces the agent to reason mechanism-first, apply per-claim provenance tags (source type + action flags), and produce role-specific implications instead of regional essays.
Tech stack
- Plain markdown skill files (
AGENTS.md, SKILL.md, runtimes/codex/SKILL.md, runtimes/claude/SKILL.md, runtimes/openclaw/SKILL.md, STATUS.md).
- Lightweight Python validator (
scripts/validate.py).
- GitHub repository with CI running the validator.
- Markdown documentation for source guide, currency watch, companion patterns, cold-start interview, practice profile, adversarial cases, and agent-eval records.
Relevance
This project demonstrates how I think about useful agent infrastructure for high-stakes regional reasoning: small reusable layers, mechanism-first contracts, honest evidence discipline, and outputs aimed at sanctions, banking, corridor, energy, and political-economy decisions — composed cleanly with a horizontal skill and a separate infrastructure layer instead of bundling everything into one repo.
Project links
Author: Vassiliy Lakhonin