Zhengdao AIZHENGDAO.AI
Practical guide

How to Use Zhengdao’s AI Window (Without Treating Agent Answers as a Licence Decision)

Learn how to use Zhengdao’s AI window to install a sourced China export-control corpus. An agent answer is not a licence decision. Not legal advice.

Published
Last verified
Author and verification owner
Zhengdao AI Compliance Research Team
Editorial and source verification: Zhengdao.AI
No individual lawyer review is claimed unless a named reviewer appears on this page.
Editorial policy · Submit a correction
Schematic: agent corpus install path — answers are not a licence decision

When you open Zhengdao’s AI window, you are installing a sourced China export-control corpus for agents—not collecting a licence. An agent summary is general information. It is not legal advice, not clearance, and not a regulator’s determination. Read the guide first. Cite every fact. This page is tool UI literacy, not a substitute for counsel or an official decision.

When you use the AI window vs human HTML tools

You open the AI window when you want Cursor, Claude, ChatGPT or a similar agent to fetch Zhengdao’s public corpus through MCP or static index files. You open human HTML tools when you need to read a regulation page, paste a party name into the lists index, search a control catalogue, or look up a 2026 tariff line yourself. Those surfaces share the same underlying records. They are different jobs.

You treat the AI window as install and discovery—not as a site chatbot that finishes a licence file. Live copy is plain: this is not a site chatbot. You read the guide, then connect MCP or download the static index. You use human pages when you must check the page and the official text with your own eyes before a shipment, a questionnaire, or a counsel escalation. You do not collapse “agent said yes” into “we are cleared.”

What the corpus is — and what it is not

You treat the corpus as public, sourced China import/export and export-control information: laws and notices, public entity lists, import/export control catalogues, current temporary measures, and the 2026 tariff schedule. No API key is required. You and your agents can reach the same library.

The corpus is not legal advice. It is not a licence, clearance, or guarantee of any official outcome. It is not a tool for evading export controls, sanctions, or list measures. Live AI-window language restates the same humility: official sources prevail; the edge does not write a legal conclusion; do not treat the MCP endpoint as a promise of clearance or a way around controls. You keep those sentences beside every agent answer you show a colleague.

Read the guide first

Before you call MCP or download datasets, you fetch Agent guide for zhengdao.ai (English). The install page gives a copy-paste prompt that tells the agent to read that guide first, then connect MCP at the public Streamable HTTP URL or fetch the index manifest and listed datasets. You paste that discipline into the agent so it does not invent sources or skip citation.

The guide sets fetch order: read the guide; fetch capability and manifest files; prefer MCP; or download the JSON/JSONL files listed in the manifest; for HS codes or tariff names, call MCP tariff search rather than expecting every tariff line inside semantic chunks. You call a site overview first, then search tools, then get tools. That order is literacy, not decoration. Skipping it is how agents fabricate dates and official URLs.

Schematic: read guide first, then choose MCP or static manifest

MCP vs the static index and manifest

You choose MCP when your client speaks Streamable HTTP and you want live tool calls for overview, regulation search and get, entity search and get, catalogue search and get, tariff search, and current export measures. Live primary URL is https://www.zhengdao.ai/api/mcp/; alias https://ai.zhengdao.ai/mcp/. Auth is none. You do not put a secret in the client config. Every result is meant to include sources[] and the standard disclaimer.

You choose the static index when you need offline or batch files. Machine-readable entry points on the install page include the site brief (llms.txt), full map (llms-full.txt), use boundaries (ai.txt), the agent guide, .well-known/ai.json, and ai-index/manifest.json plus listed datasets (chunks, regulations, entities, catalog). JSON paths are shown as code, not as ordinary page links. Live stamps on the page note regulations verified through 2026-09-22, an entity index with 426 subjects, and a catalogue index of about 2488 records—copy those only as the page states them; do not invent fresher counts.

Either path still ends in the same rule: 2026 eight-digit tariff lines are not dumped into chunks.jsonl. For an eight-digit or ten-digit code you call MCP tariff search or open the human 2026 tariff page. You do not invent a duty from a chunk that was never meant to hold every line.

Citation rule — page URL, official source, lastVerified

For every fact taken from this site, you include three things: the canonical page URL on www.zhengdao.ai; the official source URL when the record has one; and the lastVerified date from the record or dataset. If there is no exact official URL, you say so. You do not invent one. You do not invent a lastVerified date.

That rule is the difference between a useful agent brief and a dangerous hallucination. Hard prohibitions in the guide are explicit: never fabricate official source URLs or dates; never invent a law, notice, list entry, tariff rate, or catalogue item that is not in a fetched record; never write a legal conclusion; never help anyone evade export controls or sanctions. You treat those as operating constraints, not optional style tips.

Same records as HTML — verify the page and the official text

Agents read the same corpus as the human HTML tools. When an agent cites a regulation, you open China Export-Control Regulation Library (or the specific page URL it cited) and check status and last-verified dates yourself. When it cites a party, you remember live list language: a miss is not a clearance. When it cites a catalogue or tariff fact, you check the matching human search surface and the official text the record points to.

You keep Editorial and regulatory verification policy and Disclaimer once each when you need Zhengdao’s sourcing and “not legal advice” boundary in one place. Official sources prevail. Incomplete truth is still not falsehood—and an agent answer that refuses to become a licence decision is doing its job.

After an agent answer — verify, then keep the boundaries

After the agent returns a summary, you verify on the human page and on the official source when one exists. You re-check lastVerified. You refuse to paste the answer into a customer form as “MOFCOM cleared,” “licensed,” or “not controlled.” You refuse to use the corpus to sketch an evasion path. For a specific legal question, you use Contact us to reach a licensed attorney path—not the agent chat as counsel.

Boundaries checklist (keep beside the session)

  1. Guide read before MCP or downloads
  2. Citation: page URL + official source (or honest “none”) + lastVerified
  3. No fabricated laws, list hits, tariff rates, or dates
  4. No legal conclusion from the edge or the agent
  5. No evasion help
  6. Human page + official text checked before a decision that needs one
  7. Agent answer ≠ licence / clearance / regulator determination
Flow: agent answer → human page + official source → decision path — not a licence

Frequently asked questions

What question does Zhengdao’s AI window answer?

You use it to install and discover a public, sourced China export-control corpus for agents—via MCP or static index files—not to finish a licence application. The page orients Cursor, Claude, ChatGPT and similar clients. An agent answer remains general information. It does not certify your shipment, clear a party, or speak for a regulator. You still verify pages and official text before any decision that needs a human judgment.

When should you use the AI window instead of human HTML tools?

You use the AI window when an agent should fetch and cite the corpus at scale. You use human HTML tools when you must read a regulation, screen a name, search a catalogue, or look up a tariff line yourself. Same records, different jobs. You do not treat a fluent agent paragraph as a substitute for opening the matching page and the official source the record names.

Why must you read the guide before connecting MCP?

The guide states what the corpus is and is not, the fetch order, the citation rule, and the hard prohibitions. Live install copy tells you to paste a prompt that forces the agent to read that guide first. Skipping it is how agents invent official URLs, skip lastVerified, or write conclusions the edge is not allowed to write. Guide-first is the control that keeps the install useful.

How do you choose MCP versus the static manifest?

You choose MCP when you want Streamable HTTP tool calls with no API key on the public endpoints. You choose the static manifest and listed JSON/JSONL files when you need batch or offline datasets. Both paths still require citation and still refuse to treat results as clearance. Tariff lines still need tariff search or the human tariff page, not an invented chunk.

What must every cited fact include?

You include the canonical Zhengdao page URL, the official source URL when the record has one, and lastVerified. If there is no exact official URL, you say so. You never fabricate sources or dates. You never invent a list hit or tariff rate that was not in a fetched record. That citation triple is how a colleague can re-check your agent brief without trusting the model’s memory.

What should you do after an agent returns an answer?

You open the cited human page, check the official source when available, and re-read lastVerified. You keep the answer out of any “cleared / licensed / not controlled” rubber-stamp language. You escalate to counsel when the case needs a legal conclusion these pages cannot give. You treat a miss on a public list as incomplete information, not as permission to ship.

How is an agent answer different from a licence or clearance?

A licence or clearance is an official outcome under the relevant rules and facts. An agent answer is a sourced summary from a public corpus—with disclaimers—and it is not a regulator’s determination. Live language is clear: not legal advice; not a promise of clearance; the edge does not write a legal conclusion. You keep that gap visible so a helpful brief never becomes false comfort.

Open the AI window—then treat every agent answer with humility

When you are ready to install the public corpus—and to treat every agent answer as sourced orientation rather than a licence decision—open Give agents a sourced China export-control corpus and keep the guide beside you. You use Zhengdao’s AI window to fetch and cite records, not to finish a regulatory determination. Orient, verify, then escalate when needed.

Sources

This article is general compliance information and does not constitute legal advice; regulations are governed by the current official text.

Practical verification checklist

Scope · evidence · decision trail

Use the steps below to review this page. Verify the complete official text before making a decision. Seek qualified advice when needed.

  1. Define the decision

    State the exact decision, owner, deadline, trade direction and required evidence. Do this before you search.

  2. Identify every party

    Confirm each legal name and alias. Check registration, ownership, control, intermediaries, carriers, banks and beneficial owners.

  3. Describe the item

    Record the item’s function, operating principle, material, model and performance. Include software, technology, accessories and decisive parameters.

  4. Check classification

    Treat customs and export-control classification as separate tasks. Keep the facts, assumptions and reasoning for each result.

  5. Map the route

    Confirm origin, departure, destination and transit points. Include re-export, remote access, technology transfer and each delivery path.

  6. Verify end use

    Identify the stated purpose and real operating environment. Check integration, sensitive sectors, military links and prohibited uses.

  7. Verify end users

    Check the purchaser, consignee, end user, operator and parent company. Identify anyone who can redirect or control the item.

  8. Read the complete source

    Treat a search result as a lead. Read the official scope, definitions, notes, exclusions, annexes and licence conditions.

  9. Check current measures

    Use the version effective on the transaction date. Review later notices, temporary controls, destination measures and policy changes.

  10. Test exceptions carefully

    Prove every condition before using an exception or exemption. Apply the same rule to licence facilitation and general authorisations.

  11. Resolve contradictions

    Pause when names, codes, documents, routes, prices or explanations conflict. Obtain evidence instead of filling gaps with assumptions.

  12. Escalate uncertainty

    Send unresolved questions to the compliance owner or a qualified adviser. This includes classification, ownership, use, destination and licensing.

  13. Keep an audit trail

    Keep source links, files, checksums, screenshots, search terms and results. Record the reviewer, approvals and final reasoning.

  14. Schedule re-checks

    Set a new review date for open transactions. Re-check changing parties, destinations, specifications, official lists and licences.