Proof-of-Silence™ v1.1.2 · Python · offline
Seatbelts for AI agents.
A cryptographic accountability layer. Before an agent acts, an action gate runs five checks. Every decision, including the decision not to act, is hashed, signed and chained, and anyone can verify it offline.
$ pip install pydantic cryptography pytest $ python3 -m pytest -q 548 passed in 32.76s $ python3 examples/gate_demo.py demo-out 1. Hidden instruction in the attachment PERMISSION FAIL NO_VALID_GRANT VERDICT REJECT 2. The legitimate reply, under the lead's grant VERDICT ACT_NOW 3. The same reply request, replayed VERDICT REJECT Chain: VERIFIED — 4 records
- 548tests, incl. adversarial
- 21conformance vectors
- 5checks before every action
- 0 / 1 / 2verifier exit codes
- 0network calls to verify
The action gate
Five checks. All pass, or the agent does not act.
| EVIDENCE | Is there evidence for this decision, and is it recorded by hash? |
|---|---|
| PERMISSION | Is there a grant signed by a key with the grant role? Never by the agent itself. |
| POLICY | Is the action allowed, from an allowed instruction origin, within parameter caps? |
| FRESHNESS | Is the evidence recent enough for this decision? |
| REPLAY | Has this exact request already been decided? |
Permission, policy or replay fails → REJECT. Only evidence or freshness missing → WAIT. All pass → ACT_NOW. Every verdict is written to the chain. Text in an email attachment can be evidence, but it can never carry a signed grant.
Editions
Same code. The licence is the difference.
| Lite | Standard | Commercial | |
|---|---|---|---|
| Price | Free | €49 | €1,497 |
| What you get | Checklist and templates (PDF) | Full software package | Full software package |
| Action gate, Agent Firewall, Agent Ear | — | ✓ | ✓ |
| Offline CLI verifier | — | ✓ | ✓ |
| 548 tests · 21 conformance vectors | — | ✓ | ✓ |
| Use in your own organisation, incl. production | Materials only | ✓ | ✓ |
| Embed in a product, service or SDK you sell | — | — | ✓ |
| Hosted or managed service for third parties | — | — | ✓ |
| Redistribute and sublicense within your product | — | — | ✓ |
| Updates for 12 months | — | 1.1.x patches | 1.1.x and 1.2.x |
| Email support for 12 months | — | — | Best effort, no SLA |
| Download | Buy Standard | Commercial licence |
One licence per organisation, one-time payment, no subscription. The LICENSE file in the download is the binding text. Questions about your deployment?
See it work before you buy
A two-minute video: from download to 548 passed.
Check the download, install three libraries, run the tests, see the verifier's three outcomes, and watch the gate reject a hidden instruction in an email attachment.
Or skip the install: run the real example chains from the package (valid, tampered, truncated and empty) through the verifier in your browser.
Trust boundary
What a VERIFIED result means, and what it does not.
Establishes, for the presented records
- Record hashes recompute from the canonical data
- Ed25519 signatures resolve to an authorised key in the supplied registry
- Sequence and previous-hash links are continuous
- Decisions satisfy the chain's own policy
- Each ACT_NOW is backed by a genuine, unspent grant and the gate policy, when they are supplied
- No records missing from the end, when an independently held checkpoint is supplied
Does not establish
- That an external event happened as recorded
- That the key registry is genuine, unless obtained independently
- That an AI model's reasoning was correct
- That an action never reached the gate: your application must call it before acting
- Legal approval or a security guarantee
Supports key EU AI Act requirements around record-keeping, traceability and human oversight. Not a certification. Audit status: adversarially tested by the author and QA-tested by other AI models; no independent professional security audit yet. Release hashes and history →
Questions
Does it stop my agent, or just record what it did?
Both, if your application calls the gate before acting: on REJECT or WAIT the action does not run. Every verdict is recorded and signed.
Does it need the internet or your server?
No. It runs and verifies fully offline. Nothing is sent to neonai.shop.
Is this EU AI Act certification?
No. It supports key EU AI Act requirements around record-keeping, traceability and human oversight. It is not a certification.
Standard or Commercial?
Same code. Standard: use within your own organisation, no resale. Commercial: build it into products and services you sell.
Enterprise pilot — €4,900
Implement Proof-of-Silence™ for one real AI workflow: decision gate, evidence policy, stop and escalation paths, and an implementation report with stated limits. This is a scoped pilot, not certification or a guarantee of regulatory compliance. Contact sales@neonai.shop.