AI Behavior Attestation Layer

English · 中文说明 →

Your AI can do anything. Can you prove what it did?

An open-source, independently verifiable evidence layer for AI API calls and research priority.

Live demo: ai-attestation.com · License: MIT

---

The hook

July 2026. Elon Musk acknowledged that his AI company uploaded customers’ private code without consent.

What if he had not admitted it?

September 2026. Tristan Buckmaster and Levent Alpöge had privately worked for nearly a year on a blow-up solution to the Navier–Stokes equations, and uploaded an unpublished draft to OpenAI Codex for assistance. OpenAI then ran a large swarm of AI agents along the same obscure path and claimed the proof.

The fight is not “who is better at math.” It is:

• When did I already have this idea? There is no evidence.
• Which model did I talk to, and what did I say? The vendor’s word.
• Did I consent to training use? A memory.
• How would a journal or institution check? Screenshots.

Today, AI companies’ behavior rests on self-attestation. They can see your data, call your APIs, and make decisions that hit your business — and no independent third party can audit what they actually did.

We built this. We shipped our first public version in September 2026.

---

What we built

A third-party attestation layer that does not depend on any AI vendor.

Capability	What it does
API call attestation	Point your SDK base_url at our proxy. Every call appends a SHA-256 hash chain. Bodies are not stored.
Research priority deposit	Before you send a draft to any AI, hash it locally in the browser and write the fingerprint onto the chain.
Offline verify pack	One-click ZIP with verify.html. Works offline, on another machine, without our servers.
Public verify link	A URL a counterparty can open. Chain integrity, no login, no download.
Compliance as code	Open templates (EU AI Act, China’s Interim Measures for the Management of Generative AI Services, and more), versioned and runnable.

In one line: customers are not buying “we know AI better.” They are buying decision insurance — reconcile, replay, and let a third party verify when something goes wrong.

---

Why open source

Attestation only works if it is open.

A closed audit tool cannot prove it did not tamper with the data. Code, algorithms, and verification methods have to be public, or a third party cannot independently recompute them.

Compliance templates and verification methods are MIT. The hosted product is the batteries-included path. If you do not trust us, fork it, run it, and verify it yourself.

---

Who this is for

• Teams that ship AI features and need a paper trail when a client asks “what did the model do?”
• Researchers who want a timestamped fingerprint before a draft goes into someone else’s model
• Buyers, auditors, and counterparties who should not have to trust a vendor-console screenshot
• Not a stand-in for counsel, a notary, or a full GRC suite

---

Honest boundaries

This can support:

• A given hash existed at a given time
• An API call happened and request/response hashes were not rewritten on our chain
• The hash chain is intact

This cannot prove:

• A vendor trained on or read your content
• Your paper strictly predates someone else’s
• A substitute for counsel, a notary, or a qualified TSA (trusted timestamp authority)
• That a court will admit it

We would rather draw the line now than have you discover it in a fight.

---

Quick start

Hosted (fastest)

1. Open ai-attestation.com → Key Management → create a key → copy ata_xxxxxx.
2. Point your SDK at the proxy:

from openai import OpenAI

client = OpenAI(
    base_url="https://ai-attestation.com/v1/proxy",
    api_key="sk-your-upstream-key",  # DeepSeek / OpenAI / Anthropic / …
    default_headers={
        "X-Attest-Key": "ata_your_key",
    },
)

resp = client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "hello"}],
)

3. Back on the site → API Call Log → open a row → inspect the hash chain.
4. Call detail → Export Verify Pack → unzip → open verify.html. No network, no login, no need to trust us.

Without an upstream key, use Simulate a Call in the dashboard (POST /v1/demo/simulate).

Self-host

# API — http://127.0.0.1:8004
cd backend/app
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --host 127.0.0.1 --port 8004

# Dashboard — http://127.0.0.1:3002
cd frontend
npm install && npm run dev

Local proxy: http://127.0.0.1:8004/v1/proxy with header X-Attest-Key. Optional env: ATA_HOME, ATA_SMTP_*, ATA_DASHBOARD_URL, ATA_UPSTREAM_*. See the source and SECURITY.md — do not expose an unmodified MVP to the public internet.

cd backend && PYTHONPATH=app python -m pytest tests/ -q

---

Relation to SAIR's Open Math Model Initiative

In September 2026, Terence Tao and fellow Fields Medalists launched SAIR Foundation's Open Math Model Initiative, which states three principles:

• Data should not enter training sets without explicit researcher consent
• Model outputs must have traceable provenance
• Results must be reproducible and independently checkable

AI Behavior Attestation Layer provides the technical implementation for all three:

• Explicit consent: per-vendor training-consent records, each change appended to the evidence chain
• Traceable provenance: research priority deposit and call-to-artifact timeline
• Reproducible checks: offline verify packs and public verify links

We are not affiliated with SAIR and do not replace its governance framework — we provide the "verifiable evidence" layer beneath it.

FAQ

How is this different from LLMOps tools? LLMOps helps you debug and monitor your app. We are a vendor-independent third party: not OpenAI’s console, not Anthropic’s, not anyone’s. The hash chain can be recomputed offline by anyone. LLMOps is for your team. This is for customers, auditors, and counterparties.

Do you store our prompts and completions? No. API attestation stores hashes, not bodies. Research deposit is stricter: the browser hashes locally; only a 64-character fingerprint is uploaded.

Can a customer really verify without you? Yes. The ZIP includes verify.html. Offline, another laptop, no our servers. Public verify links need no download.

Is this legally binding? We produce technical evidence and a timeline — not a legal opinion, and not a substitute for a notary. See Honest boundaries.

Can you prove OpenAI trained on our data? No, and we will not say that. We can show that you possessed some content at a point in time (fingerprint + timestamp), and that a call to a vendor happened (hash chain). Whether the vendor trained on it is a question for their logs and legal process.

If it is open source, why not just fork? You can. Templates and verification are MIT. Hosting is the “give a customer a link” path: no server, no verify page, no share-token ops. Fork to learn; use hosted when a counterparty needs a URL today.

Pricing? Open source is free. Hosted is trial / subscription. We are in a pilot.

---

Architecture

Your SDK
   │
   ▼
/v1/proxy  ──►  upstream (OpenAI / Anthropic / DeepSeek / Zhipu / …)
   │
   ├─ hash only; no bodies
   ├─ SHA-256 chain: prev_hash → chain_hash
   └─ unified evidence chain
         │
         ▼
    dashboard · compliance · offline pack · public verify URL

• Backend: FastAPI + SQLite
• Frontend: Next.js 14
• Crypto: SHA-256 chain, local timestamp; optional OpenTimestamps / on-chain anchor
• License: MIT

---

Contributing

See CONTRIBUTING.md. We especially want:

• Compliance YAML templates
• Vendor adapters
• Security review of the verify path

---

Contact

• Issues: GitHub Issues
• Demo: ai-attestation.com
• ICP: 粤ICP备2026110101号 · 粤公网安备44195302000214号

Vulnerabilities: email, do not file a public issue. See SECURITY.md.

License

MIT © 2026 liao-zhizhi