Payments regulation never stops moving. A research agent watches the regulators and the rails, drafts the changes as pull requests, and proposes them — but nothing reaches the site until a payment domain-expert partner reviews and approves it. The machine drafts; the human decides.
A static knowledge base in a fast-moving regulated industry is wrong within weeks. PayShap adds a feature, MiCA limits a token, SAMA issues a directive, PCI revises a standard, PASA winds down. Keeping a payments tree accurate by hand does not scale.
The research & regulatory-watch agent is the editorial engine that solves this. It continuously monitors authoritative sources, detects material changes, and drafts proposed updates to the affected leaves as pull requests. It never publishes on its own. Every change is moderated by a payment domain-expert partner before it merges — the same review-and-stamp flow that produced the leaf you are reading.
The agent is the author. The partner is the editor-in-chief. Nothing ships unread.
The loop runs end to end with a human gate that cannot be skipped:
// The editorial loop — the human gate is structural, not optional 1. WATCH monitor SARB, PayInc, PCI SSC, FATF, PSD3/EBA, MiCA/ESMA, SAMA, CBUAE, scheme bulletins — on a schedule 2. DETECT diff against what each leaf currently says 3. DRAFT write the proposed leaf edit + cite the primary source 4. PROPOSE open a Pull Request on 2nth-ai/pay-2nth — preview auto-deploys to a per-branch URL 5. MODERATE CODEOWNERS requires a domain-expert partner review the expert reads the preview, edits or approves 6. SHIP on approval + merge → production deploy the leaf gets a fresh "Reviewed by / Last reviewed" stamp
Model choice is a tiering decision, not a single commitment. The default runtime is Cloudflare Workers AI — edge models that run cheaply and close to the request, used for the always-on work: monitoring, summarisation, public Q&A, and the training tutors.
Registered partners get Gemini or Claude via API for deep, ad-hoc tasks taken on the site — long-context regulatory reasoning, drafting a complex change, diligence work. This maps onto the access tiers: open/reference is served by edge models; member and partner depth unlocks frontier APIs.
| Tier | Default model | Used for |
|---|---|---|
| Open | Cloudflare Workers AI (edge) | Monitoring, summaries, public Q&A, tutoring |
| Member | Edge + selective frontier | Decision-content agents, training assessment |
| Partner | Gemini / Claude via API | Deep ad-hoc reasoning, drafting, diligence |
The agent never calls a provider SDK directly. Calls route through a model gateway, so “edge vs frontier” stays a configuration decision, not a code rewrite. This is the same portability stance as the 2nth-ai/agent-platform control plane this references — open weights through frontier, never locked to one vendor.
An unmoderated AI that edits a payments compliance page is a liability, not a feature. The value of this surface to an executive, a PSP or a consultant is that a named domain expert stands behind every Live leaf — with a review date and a primary source. The agent makes that sustainable by doing the watching and drafting; the partner makes it trustworthy by deciding what is true.
This is the difference between “AI-generated” and AI-drafted, expert-approved. The first is noise. The second is a maintained, citable knowledge base that an agent can also load as context.