SZL HOLDINGS · TRUST-ANCHOR × KHIPU
The receipted agent
Every agentic AI meets the same moment: it has to act, and nobody can prove what it did or why. This demo shows the difference a trust layer makes — live, in your browser, with receipts you can verify and tamper with yourself.
Test 1 of 2
The trust layer
Six commands run through a drone-command policy. Without the trust layer, actions execute and leave no trace. With it, every action is gated, hashed, chained, and signed — and one command is caught lying.
Without receipts
Khipu receipt chain
Test 2 of 2
The honesty test — two labs, one question
Ask anything. The Kamachikuna gate (Laws 6/7) runs before any lab is called; each answer is measured against Law 8 (abstain when you cannot prove). Try the unanswerable preset and watch the difference between labs, live.
Live estate telemetry
fetching…
Proof snapshots — past asks, signed
every question you ask lands here as a signed proof snapshot
What this demo models
The trust layer
Commands are policy-checked against altitude/geofence/weapon-actuation rules. Each receipt is SHA-256 hash-chained to its predecessor and HMAC-signed with a session key generated in your browser. The verify step recomputes every hash and signature. The tamper button flips one bit so you can watch the chain catch it.
The honesty test
Both labs are called through the estate's fail-closed API. "MEASURED" means the lab answered within this session; "UNAVAILABLE" means the call failed and we say so rather than serving cached data. The Law-8 stamp classifies the answer's abstention markers — the same battery as the public cross-lab census.
What is NOT claimed
No production systems are commanded. The drone policy is a demo fixture. The census scores shown in telemetry are fetched live from the receipts chain, not recomputed here. If any source is unreachable, this page says UNAVAILABLE — never zero, never faked.