AiXin Academy · EN / 中文
Learn governed AI twins — and prove it to someone else
20 modules across 4 tracks. Every module states what you can do afterwards, checks your understanding instantly, and ends in a hands-on exercise. Every claim is labelled Live, Sandbox or Planned — if you can't open a URL for it, it's Planned.
Can / cannot do
The honest capability matrix, every row pointing at the code that enforces it.
Twin templates
Seven copyable manifests and tutorials, each with a buyer and a monetisation model.
Time-and-money map
What to delegate first, hours back per month, and payback.
Track 1 — Foundation 入门
Non-technical · phone only · 90 minutes
What AiXin is
AiXin is a trust layer for agentic AI, not another chatbot. It does not make the model smarter — it makes the model's actions accountable.
Master Twin + Specialist Twins
One Master Twin per account orchestrates a team of Specialist Twins. Orchestration and execution are deliberately separated.
The SIP pipeline
Intent → validation → Decision Card → execution → signed receipt anchored on BSC Testnet. Five stages, one human checkpoint.
LAB — your first governed action
Phone only, no VPN: hatch a twin, install a skill, approve one card, reject one, then open the receipt on BscScan.
Track 2 — Builder 构建
Developers · laptop or GPU box · Half day
Skills & SKILL.md manifests
A skill declares its scope, visibility (public/private), pricing (free/paid), status (dev/live) and version in a manifest.
Dynamic tool registry + scope enforcement (W1/W2)
The model searches a tool registry instead of receiving every tool. Scope enforcement resolves whether a twin may invoke a tool at all.
Adapters and execution honesty
Telegram, Gmail, signed Webhooks (HMAC-SHA256), WeChat and BSC. With no live adapter, the run halts as BLOCKED — it never pretends.
Run safety: budgets & kill switch (W4)
Step, token and wall-clock budgets per run, a monthly workspace cap, and an operator kill switch. Every run records its stop reason.
Prompt traces, replay & dsh
Traces log everything the model could see. Deterministic replay is NOT built yet, and the dsh runtime bridge is planned for self-hosted boxes only.
Self-hosting in China
Ubuntu or Windows 11 GPU box, Qwen via Ollama, local database, no cross-border egress required. Same build, configurable model source.
Track 3 — Community & Investor
Community leads, partners, investors · 60 minutes
Why governance is the moat
Model capability commoditises. Verifiable governance — evidence a third party can check — compounds.
Marketplace economics (W5)
Twin Bundles, creator earnings and a platform fee. Planned after mid-September — never present it as live.
Case study — BangBang built on AiXin
BangBang is a tenant application built on the AiXin platform — multi-tenancy, content safety, a WeChat channel and a tenant API — not an external app calling AiXin.
The three live demo moments & the ask
Never use slides for: approving a Decision Card, signing a rejection, and a stranger verifying a receipt. Then make one concrete ask.
Track 4 — Build & monetise 构建与变现
One-person companies, small business owners, skill creators · Half day
What your twins can and cannot do
The honest capability matrix. A twin can read, reason, draft, delegate and execute through a connected adapter — it cannot browse the open web, move money without an approved Decision Card, or run without a live adapter.
Create your first twin in 30 minutes
One worked example end to end: hatch a Specialist, write a skill with deterministic rules, connect one adapter, delegate, approve, and open the signed receipt.
The OPC time-and-money map
Which tasks a one-person company or small business should delegate first, how many hours each returns per month, and the payback maths — including the tasks you should not delegate yet.
Ten ways a twin is monetised
Monetisation is broader than selling: cost avoided internally, one-off build fees, subscriptions, usage-based pricing and marketplace bundles. Some models are live today; in-app paid commerce is planned (W5).
Best practices and anti-patterns
What separates a skill people trust from one they uninstall: narrow scope, deterministic rules, real evidence on the Decision Card, budgets, versioning, and an honest failure path.
Why AiXin, commercially
OpenClaw, Hermes and DeepSeek Harness optimise for capability and composability. AiXin optimises for evidence a third party will accept — which is what makes a twin sellable into a business.
Glossary (use these words)
- Master Twin 主孪生
- Your single orchestrator; reads intent, delegates, never executes blind.
- Specialist Twin 专家孪生
- A purpose-built worker twin with a limited set of skills.
- Skill 技能
- A capability unit declared by a SKILL.md manifest.
- Adapter 外部工具连接
- The connection to a real outside system: Telegram, Gmail, Webhook, WeChat, chain.
- SIP (Signal Intent Protocol) 信号意图协议 SIP
- Intent → validation → approval → execution → receipt.
- Intent signal 意图信号
- The strict JSON the model proposes. A signal is not an action.
- Decision Card 决策卡
- The human approval surface, with an evidence panel.
- Signed receipt 签名回执
- An Ed25519-signed record of what actually happened.
- Anchor 上链锚定
- Writing the receipt hash to BSC Testnet so outsiders can verify it.
- Run budget 运行预算
- Step, token, time and monthly limits applied to a run.
- Kill switch 紧急停止开关
- An operator pause that refuses new runs pre-flight.
- Prompt trace 提示词轨迹
- The log of everything the model could see in a run.
- Tool registry 工具注册表
- The searchable tool catalogue offered to the model.
- Scope enforcement 范围强制
- A twin may only invoke tools inside its assigned skills.
- Capability matrix 能力清单
- The can / cannot / refuses-by-design list for twins.
- Draft (not executed) 草稿（未执行）
- Output produced when no live adapter could perform the action. Nothing left the platform.
- Monetisation model 变现模式
- How a twin turns into value: cost avoided, build fee, subscription, usage-based, or marketplace bundle.
- Cost avoided 节省的成本
- Monetisation by removing hours from your own business — live today, no commerce needed.
- Payback 回本
- (hours saved × hourly rate) − (run cost + amortised build time).
- Twin template 孪生模板
- A copyable starting manifest plus a tutorial, buyer and monetisation model.
Objections & honest answers
Isn't this blockchain hype?
The chain does one narrow job: making a receipt hash tamper-evident to a third party. Validation, approvals and execution are ordinary server code and work without it.
Why can't the AI just do it?
It can, and that's the problem: an unaudited agent's mistake is indistinguishable from its success. We keep the model for judgement and give the action to deterministic code.
Isn't this slower than an autonomous agent?
Reversible actions are not gated at all. Gating applies to consequential actions, where a wrong move costs money or trust.
Can you prove it, not just claim it?
Yes — open a receipt's public verify link on your own device, no account needed.
What if foreign model providers are unavailable in my region?
The model source is configurable. A GPU box running Qwen via Ollama runs the same build locally (M10).
Do you store my data abroad?
Self-hosted deployments have no cross-border egress requirement; see the deployment runbooks.
Is the marketplace live?
No. Skills, manifests and installs are live; paid commerce is planned (W5). Any trainer who says otherwise is wrong.
Can I replay a run?
Not yet. Prompt traces are recorded — deterministic replay is planned, and trainers must not demo it.
Can a twin make money for me?
Say it precisely: a twin is monetised. Today that means hours removed from your own business (cost avoided), or work you deliver to a client and invoice outside the platform, with the signed receipt as proof of delivery. In-app paid commerce is planned (W5) — do not sell it as live.
What can a twin definitely not do?
Browse the open web at will, act without a connected adapter, move money without an approved Decision Card, or run without step/token/time budgets. Each of those is a hard refusal in code, not a policy promise.
Certification: AiXin Skill Creator
- A live skill with a valid SKILL.md manifest
- The skill installed AND assigned to a Specialist
- An approved Decision Card with a real evidence panel
- A signed rejection
- A receipt verified on BscScan by someone else
- Optional: a skill shipped from a twin template, with a named buyer and price
Did it work? (per 20-person workshop)
| Learners who hatched a twin | ≥ 18 / 20 |
| Learners who approved AND rejected a card | ≥ 15 / 20 |
| Receipts verified by a different person | ≥ 10 |
| Skills published (Track 2) | ≥ 5 |
| Certifications issued within 2 weeks | ≥ 3 |
| Modules where >30% failed the knowledge check | 0 |
Review these, then change the curriculum — not the audience.