Case Study: Gulf + Middle East Hybrid Intelligence Skill

Key facts

Role: Builder and maintainer.
What it is: A vertical specialist skill for AI agents working on Iran sanctions exposure, GCC correspondent banking and trade finance, oil and LNG markets and OPEC+ behaviour, maritime chokepoint risk (Hormuz, Bab-el-Mandeb, Red Sea), sovereign-wealth deployment, and Iranian proxy-network exposure.
Form: Plain markdown skill files that attach to any capable agent — canonical SKILL.md plus Claude and Codex runtime overlays. No CLI, no runtime, no hosted service.
Current state: Bar 1 and Bar 2 cleared under the repository's historical agent-integration canon; the B2.8 practitioner review is not recorded. STATUS.md carries the current position.
Boundary: Structural discipline only. Not sanctions or vessel screening, not factual verification, not advice.
Claimed usage: None. No production use, adoption, or benchmark numbers.

What it changes in the output

Generic model output on the region narrates countries and ends on "monitor the situation". It rarely traces how a designation, a chokepoint incident, or a sovereign-wealth deployment reaches bank exposure, refining margins, charter rates, or insurance premia. It also collapses Iran-state, IRGC-affiliated, and Iran-private commercial actors into one actor — the distinction most sanctions questions turn on.

The skill is a behaviour contract. It requires:

mechanism first: the transmission channel before the implication;
per-claim provenance tags — [primary], [secondary], [inference], [analyst-judgment] — plus [verify] and [stale-risk: YYYY-MM] action flags;
a currency trigger: live verification on sanctions designations, OPEC+ decisions, chokepoint events, JCPOA-track status, and vessel-specific claims;
the Iran-state / IRGC-affiliated / Iran-private distinction, enforced rather than optional;
role-based implications — compliance lead, energy buyer, shipping insurer, Gulf bank, refiner, sovereign-wealth analyst — and named trigger points to watch next.

Nine flagship examples in the repository show the contract applied, each labelled with its evidence mode. One works through a European bank's sanctions desk facing an OFAC SDN listing against UAE good standing, where the operative question is not which list wins but where the bank's exposure surface actually sits.

How it is checked

scripts/validate.py runs in CI: required phrases, forbidden determinative claims, evidence-mode coverage. It checks structure, never whether a statement is true.

Externally checkable claims leave a memo through the evidence-packet handoff — stable claim IDs, declared source IDs, caller-supplied source text, optional verbatim quotes. Scenarios, assumptions, inferences, and judgments stay in the memo. The synthetic packet passes both the Agenda Intelligence MD strict CLI and the repository's own CI validator, which establishes interface compatibility, not factual truth.

Adversarial, agent-eval, and skill-improvement cases are recorded in the repository. They are self-scored structural checks under one model and one prompt run — not a benchmark, not accuracy measurement, not practitioner validation. The repository also states what the skill has not been tested on: no labelled accuracy dataset, no multi-agent trials, no cross-model regression tracking, no live-source automation, limited Arabic- and Farsi-language source coverage, and no vessel-tracking or AIS data.

What it is not

Not sanctions screening, vessel screening, or maritime due diligence.
Not a factuality verifier or live source retriever.
Not a vessel-tracking or AIS-data product.
Not legal, compliance, AML, investment, security, or operational advice.
Not an agent framework, CLI, MCP server, or validation platform.
Not a generic strategic-memo tool — that is Global Think Tank Analyst.
Not a replacement for a human analyst, counsel, or sanctions-desk review.

Portfolio context

This is one of two vertical specialists in a four-repository portfolio built to compose. Region depth lives here; the reasoning method lives in Global Think Tank Analyst; Central Asia and Caspian depth lives in its own vertical, referenced when a flow crosses both regions, such as Iran–Central Asia–Russia routing; the deterministic packet check lives in Agenda Intelligence MD. No repository restates its neighbour.