An agent saying “I did it” is not proof that it was done.
UrvaIQ sits beside your agents: it authorizes every consequential action against versioned policy, executes it through governed connectors, then independently re-reads the authoritative system — and hands you a signed, tamper-evident receipt.
urvaiq.com/console — demo tenant · seeded fixtures
Registry
- finance-agent · active
- invoice-bot · active
- vendor-bot · suspended
Verification
- VERIFIED418
- CONTRADICTED7
- INDETERMINATE3
- PENDING2
- UNVERIFIED0
run-r8f3a4 · INV-2026-00421
Post invoice → ERP “Atlas”
- Action requested · ap-dev-finance · post_invoice · INV-2026-00421
- Inspect · no secrets · no injection signals
- Authorize · REQUIRE_APPROVAL · tr-s-p2 · policy v3
- Approved · J. Rivera · payload c4f1…e921
- Executed · POST /invoice · HTTP 200 · observation #o-41 · not proof
- Verified (read-back) · ERP · separate read-only credential · amount 200.00 ≠ 100.00
receipt-8f3a4c… · Ed25519 · key-v2 ·
sha256:1f9c…e921 · supersedes: none
append-only · hash-chained
The case that proves the category
Policy said yes. The payload matched. The ERP disagreed.
The agent reported the invoice posted. Our independent read found 200.00 against a contract expected of 100.00 — and an unattended control plane would have believed it.
Step 1 — the claim
finance-agent reports: posted to ERP Atlas, no exceptions. The API answered HTTP 200 — an observation, not proof.
claim · not evidence
Step 2 — the read-back
A different runtime, different credentials, re-reads the invoice: amount 200.00 ≠ 100.00 expected. Conflict surfaced, never silently resolved.
observation · read-only
Step 3 — the receipt
CONTRADICTED is recorded with check results, an exception case opens for a human, and a signed receipt lands on the chain. Nobody is called at 3 a.m. for a partial success.
receipt · append-only
The six-step loop
Enforced before. Observed after. Proved at the end.
One continuous loop for every consequential action — posting an invoice, changing vendor bank details, refunding a customer.
01Identify
Which agent, which model and prompt version, which tenant — and whose delegated authority it is using.
02Inspect
Secrets, personal data and injected instructions are checked before the tool call runs — never by the agent itself.
03Authorize
A versioned policy decides: allow, deny, require a named human, or cannot decide. Fail closed, always.
04Execute
The approved action runs through a governed connector. The ERP's reply is an observation — not proof.
05Verify
A different service with different, read-only credentials re-reads the authoritative system and compares field by field.
06Prove
An Ed25519-signed receipt lands in an append-only, hash-chained audit trail. Webhooks fire; dashboards update.
Fail closed, in plain words
If we cannot verify it, we will not certify it.
Technical failures are never converted into business verdicts. A signing-key outage issues no receipt. A missing or unevaluated check is never VERIFIED.
The truth table
Five states. No sixth hiding place.
- VERIFIEDexpected and observed agree · a receipt with check results · billable unit 1.0
- CONTRADICTEDobserved disagrees · exception case opens · billable unit 1.0 · reviewer notified
- INDETERMINATEsources conflict or evidence partial · never a business verdict · billable unit 0.5
- PENDINGnot yet read back · re-reads bounded · never billed · no receipt yet
- VERIFICATION_FAILEDcannot observe · no verdict on the business action · no receipt issued · never billed
The refusals
What never happens
- A policy timeout waves the action through
- A missing receipt is silently marked complete
- A down connector becomes a CONTRADICTED verdict
- An expired approval authorizes a changed payload
- A self-approved SoD violation becomes law
# refusals are features
No single identity can break it
Separate runtimes. Separate credentials. One audit chain.
isolated runtime
Policy decision point
OPA, versioned. ALLOW / DENY / REQUIRE_APPROVAL / REQUIRE_ADDITIONAL_EVIDENCE / INDETERMINATE. Fail closed on outage or malformed answers.
read-only identity
Verification runtime
Egress allowlist, separate credential. Re-reads the system field by field; conflicts surfaced, never silently resolved.
no audit access
Receipt signer
Ed25519 over RFC 8785. Keys in KMS, published verification key set. It signs immutable receipts — or refuses.
Compliance — read the fine print
Control packs are configuration. They are not certification.
UrvaIQ maps control packs to the EU AI Act, NIST AI RMF, ISO 42001, SOC 2, PCI DSS and SOX ITGC — with owners, implementation status, test procedures and evidence requirements. Mappings are versioned configuration you review and operate; they do not confer certification. UrvaIQ is not currently SOC 2 or ISO 27001 certified. Any claim to the contrary is inaccurate.
What you keep
- →Evidence an auditor can inspect — without granting them raw access to your systems
- →Segregation of duties provable from the chain, not from tickets
- →Approvals bound to exact payload hashes — substitution is void
- →An exception queue with human owners, not silent failures
Honest answers
Ask this of every vendor in this category.
Does a receipt prove the agent told the truth?
No. A receipt proves the receipt is intact and came from us. It does not prove that the source system told the truth, that an accounting judgment was legally correct, or that a transaction was fraud-free. UrvaIQ verifies what the system of record contains — and refuses to verify what it cannot observe.
What happens when the ERP is unreachable?
The verification run ends VERIFICATION_FAILED — a technical failure, never a business verdict. We do not say the posting failed, and we do not say it succeeded. That refusal is the product.
How is the verification credential different from the agent's?
Completely separate. Verification uses read-only credentials and an isolated runtime with its own egress allowlist. The agent cannot see the verification identity, cannot re-read its own writes, and cannot turn a failed check into a success.
Does UrvaIQ make us compliant with the EU AI Act, SOX or PCI DSS?
No. Control packs provide configuration mappings and reviewable evidence; they are not certification and they do not make you compliant. UrvaIQ does not hold SOC 2 or ISO 27001 certification, and compliance-language on any page is configuration, not assurance.
Can agents change their own policy?
No. Every published policy version is immutable and approval-gated, and the exact version is recorded on every decision. Agents cannot modify, approve or deploy their own rules — and approval is bound to the exact payload hash, so the agent cannot change the amount after a human says yes.
What does failure matrix coverage mean for us?
We test ten dependency-failure rows — IdP, PDP, approval service, read connector, write connector, queue backlog, verification timeout, signer outage, database failover, webhook dead-letter — with automated fault injection, not runbooks. Every row fails closed while preserving accepted state.
Watch the ERP say no to the agent.
Thirty minutes, six seeded cases — correct, wrong amount, duplicate, missing, pending, unavailable — each settled with a five-state classification and a signed receipt. No theatre, no mock slides.