Trust Fabric · EvidenceIndependent, tamper-evident evidence for every AI action.
Witness signs and hash-chains what your AI did into receipts anyone can verify against a key you hold. Verified Actions adds governance on top: a policy gate, a capability check, and optional human attestation before a consequential action runs. Evidence first. Governance on top.
One witness fabric. Two primitives.
Witness is the evidence layer beneath everything. Verified Actions is governance built on top of it. The same fabric hardens from evidence, to witnessed conformance, to policy-gated action — witnessing where Attexa is beside the path and enforcing where Attexa owns the boundary — so coverage grows without a rebuild.
Signs and hash-chains tool calls, results, approvals, generated artifacts, and memory writes into independent, portable receipts. Answers: what happened?
Policy and capability checks before a consequential action runs, with optional human or domain attestation. Answers: was it allowed, and who stood behind it?
Four properties that make the evidence hold.
Independent
Evidence that does not depend on the party being governed, or on a vendor dashboard.
Tamper-evident
Hash-chained, so any change to the record is detectable by anyone who checks it.
Portable
Travels outside any system. Verify it in any language, on any machine, with no Attexa service in the loop.
Customer-owned
The evidence is yours and outlives every participant: the AI vendor, the operator, and Attexa.
Three layers of assurance, mapped to plain questions.
The default is observed. Enforcement and accountability are named tiers you reach as the stakes rise, on the same fabric. At L2, the same policy layer can prevent (enforce) or prove (sentinel) — whichever the situation allows.
The same L2 policy layer, two stances. Sentinel claims cover explicitly enrolled channels with source-bound evidence — not global coverage.
A wrapper at the action boundary, not a rebuild.
Witness attaches where your AI already executes its actions. The model does nothing differently: no retraining, no prompt changes.
Standard cryptography. A key you hold.
Witness uses standard primitives only, so the trust path is intentionally small: an open protocol, standard cryptography, and a key the customer holds. No vendor dashboard is required to verify a receipt.
{
"action": "tool.write_file",
"prev": "sha256:3f9a…c041",
"payload":"sha256:b71e…02ad",
"ts": 1781049600,
"sig": {
"alg": "ed25519",
"by": "witness:eng-01",
"val": "a7d2…99fe"
}
}
// verify against a key you hold
Three ways to adopt the fabric.
Verified Actions drops onto any surface. The deployment products each run agents on a specific footing, with Witness beneath every one.
Verified Actions
The governed-action layer for any AI surface. Policy gates, capability checks, and human attestation for consequential actions, with independently verifiable evidence. Drops onto what you already run.
SDE Workbench
A governed cloud development environment. Developers get a familiar AI coding workstation; security gets receipt-backed evidence of every model call, tool use, and developer action.
Agent Anything
A governed environment where an orchestrator builds and runs a team of agents for your organization, with the Witness layer beneath every agent so each one’s actions are signed automatically.
What it proves, and what it doesn't.
Proves what an AI did, in what order, under what policy, with what approval, independently, and lasting beyond every participant.
A small, explicit trust base: an open protocol, standard cryptography, and a key you hold. Not a blind one.
Does not expose hidden model reasoning, and does not claim enforcement unless a Verified Action policy gate is present.
Does not make you compliant; it produces the evidence your compliance program needs.
A regulator, an auditor, a customer, or a counterparty can verify a receipt against a key you hold, in any language, long after the original system or vendor is gone.
Witness turns AI action logs into independent, tamper-evident, portable evidence. Verified Actions governs which consequential actions are allowed. Together they produce the evidence your compliance program needs, owned by you.