Mandate Guarantee and receipts

Question answered: What happens when the policy engine gets it wrong? What do I get as evidence of every decision?

What delta proves, and what it doesn’t

The delta pipeline has two parts with different trust properties.

Evidence extraction: turning the agent’s proposal into structured facts — is non-deterministic. An LLM pulls relevant details (price, material, size, delivery date) from the source data and maps them into delta’s taxonomy. This step is not cryptographically proven.

Policy evaluation: checking the extracted evidence against the user’s intent — is deterministic. The policy engine returns pass or fail with any failed constraints. A zero-knowledge proof is generated here, binding the evidence, the intent, the policy, and the result. Anyone can independently verify that the evaluation was computed correctly.

That boundary — extraction is unproven, evaluation is proven — is where the Mandate Guarantee attaches.

The Mandate Guarantee

The Mandate Guarantee is a financial guarantee. If delta’s policy engine produces a false positive — the evidence indicates the proposal should fail, but the engine mistakenly passes it — delta covers the loss.

Covered: delta eats the loss in the case of two types of failures:

• Evidence Extraction Failure: The merchant’s source data shows the product is white, but the evidence extraction or taxonomy mapping incorrectly records it as black. The policy engine then correctly passes based on the extracted evidence. The user receives the wrong item.
• Policy Evaluation Failure: The merchant’s source data shows the product is white. The user asked for black. The evidence extraction correctly captures “white.” The policy engine incorrectly evaluates this as a pass. The user receives the wrong item.

Not covered: The merchant advertises a black T-shirt. The evidence correctly shows black. The policy engine correctly passes. The merchant ships a white T-shirt. This is a standard merchant fulfillment failure — outside delta’s scope.

The boundary is the evidence-to-evaluation step. If the extracted evidence was correct and the engine should have failed but didn’t, the guarantee applies. If the evidence was correct and the engine correctly passed, but the merchant behaved differently than their own data indicated, that is a merchant dispute — not a delta guarantee event.

Platforms that implement delta Mandate receive the guarantee and can pass it through to end customers as the assurance behind protected delegated spend.

Mandate execution receipts

Every evaluation produces a signed mandate execution receipt — an auditable object containing:

• The mandate (the user’s authorized intent and the policy it references)
• The proposed action (the agent’s proposal)
• The evidence that was extracted and checked
• The decision returned (pass or fail, with failed constraints)
• The policy that governed the evaluation

The receipt is backed by the zero-knowledge proof, so its contents are independently verifiable. It is the record a platform, issuer, user, or dispute resolver uses to confirm what was authorized, what was attempted, what was checked, and why the result was what it was.

Together, the guarantee and the receipt are what make agentic spend commercially usable at higher limits: the guarantee caps the platform’s downside if the engine errs, and the receipt provides the evidence trail for every transaction — whether it passed or failed.