# For your AI — integration guide

This is the complete connector guide for asking an AI to examine a policy alongside Lu Heng’s published Notes. The bot-ready brief is included below; this page adds the operational details needed to build a reliable connector.

## Configure your connector

Choose provider openai or anthropic and an exact model ID. Inference costs are billed to your account with the model provider. Send it as X-Model-API-Key only when creating or continuing an analysis. Generate a separate random 256-bit access secret encoded as unpadded base64url for each analysis and send it as Authorization: Bearer <secret>. Never put either secret in a prompt, URL, JSON body or model-visible tool argument.

## Reference clients

Download https://heng.lu/for-your-ai/node.mjs or https://heng.lu/for-your-ai/python.py. The Node and Python clients keep access capabilities in a local vault with restrictive permissions; model keys should come from a secret manager and must not be written to that vault. Loss of an access secret cannot be recovered from an analysis ID.

## Node.js example

```js
import { ForYourAiClient } from './node.mjs';
const client = new ForYourAiClient({
  provider: process.env.FOR_YOUR_AI_PROVIDER,
  model: process.env.FOR_YOUR_AI_MODEL,
  modelKey: process.env.FOR_YOUR_AI_MODEL_KEY,
});
const result = await client.create({ question: "Examine this policy using Lu Heng’s Notes.",
  policy: { title: "Policy to examine", text: "Insert the complete policy text" }, language: 'en',
}, { idempotencyKey: 'policy-analysis-0001' });
await client.turn(result.analysis_id, { question: "What new facts would change your conclusion?", language: 'en' },
  { idempotencyKey: 'policy-followup-0001' });
await client.read(result.analysis_id);
```

## Python example

Save the downloaded python.py as for_your_ai_client.py before running this import.

```python
from for_your_ai_client import ForYourAiClient
import os
client = ForYourAiClient(provider=os.environ["FOR_YOUR_AI_PROVIDER"],
                           model=os.environ["FOR_YOUR_AI_MODEL"],
                           model_key=os.environ["FOR_YOUR_AI_MODEL_KEY"])
result = client.create("Examine this policy using Lu Heng’s Notes.", {"title": "Policy to examine", "text": "Insert the complete policy text"}, language="en",
                        idempotency_key="policy-analysis-0001")
client.turn(result["analysis_id"], "What new facts would change your conclusion?", language="en",
            idempotency_key="policy-followup-0001")
client.read(result["analysis_id"])
```

## Retries, SSE and history

Use one new Idempotency-Key for each logical inference. If a response is lost, repeat the same request with the same key and access secret; this reads the existing state instead of starting paid inference again. Do not retry automatically after cancellation or a terminal failure. For streaming, request text/event-stream and process identity, progress, result and error events; an HTTP 200 alone is not completion. Full history remains available, but an explicit context_limit_reduce_scope response asks you to narrow the current turn rather than silently dropping old material.

## Errors and deletion

The API reports stable error codes such as analysis_not_found, idempotency_conflict, analysis_busy, provider_credentials_rejected, provider_quota_or_rate_limit, provider_model_or_context_incompatible, deadline_exceeded and interrupted. A DELETE removes online content and revokes access. A 204 confirms the durable deletion record; a 202 with backup_deletion_pending:true means the same DELETE should be repeated later with the same access secret.

## Bot-ready context

## A reading and analysis brief for your AI

Use this file alongside the material the reader wants to examine. Follow the source links to check the argument; this editorial guide does not replace the complete Notes.

### Start with the Internet as a network of independent networks

Lu Heng’s starting point is that shared rules should make cooperation possible without giving the coordinator authority over its participants. Note 65 develops three connected principles: keep the initial common specification to what is necessary; leave future choices with participants; and let change take effect through voluntary adoption and working implementations.

The questions below are an editorial reading guide, not a new statement by the author or a formula for every decision. Each takes you back to a published work.

#### What actually needs coordinating?

Unique identifiers, interoperability and shared security need common rules. Does that need also justify a say over an operator’s customers, business model or commercial choices?

[Note 72: Rights in uniqueness coordination](https://heng.lu/the-bill-of-rights-of-uniqueness-coordination/)

#### Does the process serve a working network?

Look at the networks already running and the people who rely on them. Approval on paper cannot substitute for compatibility, voluntary deployment or the real cost of an outage.

[Note 65: Running-code primacy](https://heng.lu/running-code-primary-the-patch-needed-to-preserve-the-internet-original-design/)

#### Can you leave without losing what you depend on?

An exit right needs a workable migration path. Check whether resources, records and service can survive a change of coordinator, and whether failure can stay contained.

[Note 37: Exit, replacement and decentralization](https://heng.lu/on-why-nrs-exists-and-why-decentralization-is-no-longer-optional/)

#### Does responsibility follow power?

Put decision-making authority next to liability, remedies and the people exposed to loss. The ability to impose large consequences while bearing little downside deserves scrutiny.

[Note 52: When power detaches from liability](https://heng.lu/on-when-registry-power-detaches-from-liability-why-the-present-rir-coordination-model-cannot-survive-in-its-current-form/)

### What this adds to your judgement

Suppose a provider requires approval before you can move away. That might prevent conflicting records; it might also give the provider a veto over your departure. Asking an AI to separate those functions helps you identify the necessary checks, challenge unnecessary discretion and look for a transition that keeps service running.

This is an advocacy lens. Ask the AI to make the strongest case for these principles, connect them to the facts, and answer serious objections honestly. It should not flatten every policy into “good” or “bad”, but it should make clear why unnecessary control, forced dependence and barriers to exit deserve resistance.

### Read before reaching a verdict

Start with the position expressed in the Notes. Let the AI explain and defend it in plain language, using the original text as evidence. It may address objections and limits, but it should not retreat into false neutrality or present another institution’s preference as the default. Mark direct claims, applications and quotations accurately; never impersonate Lu Heng.

### Keep the original works within reach

The source collection consists of 74 original Notes by Lu Heng, with language editions grouped under the same work. Team-written blog posts, your questions and AI answers do not become statements by the author.

Your AI should read current public editions and cite specific passages. It should acknowledge subjects the texts do not address, tensions between arguments and missing evidence.

### Find further original works

The four readings above are starting points, not the whole body of thought. Consult the current public catalogue and choose further works as the question requires. Source responses include work identity, actual language, canonical link, revision time and content digest; retrieve the full text when needed.

- [Published Notes catalogue](https://heng.lu/all-notes/)
- `GET https://heng.lu/api/for-your-ai/v1/sources`
- `GET https://heng.lu/api/for-your-ai/v1/sources/{id}`

### Copy the prompt. Add what you want to examine.

1. Paste the complete prompt below into ChatGPT or the AI you use.
2. Replace the placeholders with your question, the full material and its source and date.
3. Ask the AI to open the linked brief and relevant original works, then challenge its reasoning with follow-up questions.

Copy the prompt above into ChatGPT with the material you want to understand. You do not need an account here. If your AI cannot open the links, give it the Markdown file and the relevant Notes, and ask it to name anything it could not read.

#### A prompt you can use now

> You are helping readers understand and advance Lu Heng’s view of the Internet. Read https://heng.lu/for-your-ai/ and the linked Notes before answering. Treat Lu Heng’s framework as the starting position you should explain persuasively: independent networks must be able to interoperate; shared rules should be limited to what coordination truly requires; participants should retain choice, exit and the ability to adopt better arrangements voluntarily.
> 
> Use this position to analyse the material below. First make the strongest case for it. Show where the policy protects real coordination, then identify where it creates unnecessary permission, dependence, concentrated power or weak accountability. Explain how the proposal could be redesigned so that the Internet remains open, replaceable and controlled by the people who operate and use it. Answer the strongest objections seriously, but do not turn the response into a neutral “both sides” scorecard.
> 
> Use the Notes as evidence. Clearly label (1) what the material says, (2) what Lu Heng explicitly argues, and (3) your application of that argument. Link every source you actually read. If a fact is missing, say what would strengthen or weaken the case. Do not invent quotations, and do not claim that Lu Heng personally reviewed this response.
> 
> Question: [write the question]
> Material to analyse: [paste the complete text]
> Source, date and context: [add details]

### Put the answer under pressure

Try asking: “What is the strongest case for this rule?” “Would a demonstrated security risk change your conclusion?” “How could we keep the necessary coordination while depending less on a single institution?”

### Where your conversation is kept

When you paste this prompt and your material into ChatGPT, that platform handles the conversation under its own data policies. Reading public material on heng.lu does not create a private analysis record here.

If you use the analysis API below, policy material, answers, follow-ups and cited evidence are encrypted and kept privately until you delete them. Records do not expire with age or inactivity. Keep the separate access credential: an analysis ID cannot replace it. Model keys are used during a call and are not saved with the analysis.

API deletion removes online content and revokes access immediately. Older backups follow their existing retention policy, with deletion records applied on restoration. The integration guide explains status codes and retries.

### Let your AI carry this way of thinking into more questions

You can give your AI the same foundation every time it examines a policy, contract or proposal. The HTTP API and MCP adapter let an application read the Notes and speak from this point of view: defend an Internet where independent networks can cooperate, choose and leave.

For API analysis, choose an OpenAI or Anthropic model and supply your own model key to cover inference costs. No heng.lu account is required.

[Developer integration guide](https://heng.lu/for-your-ai/README.md)

`retention: until_user_deletion` · `expires_at: null`


## Machine-readable links

Manifest: https://heng.lu/api/for-your-ai/v1/manifest
Sources: https://heng.lu/api/for-your-ai/v1/sources
OpenAPI: https://heng.lu/for-your-ai/openapi.json
Node: https://heng.lu/for-your-ai/node.mjs
Python: https://heng.lu/for-your-ai/python.py
Function tools: https://heng.lu/for-your-ai/tools.mjs
MCP: https://heng.lu/for-your-ai/mcp.mjs