Decision receipt
Records what the rulebook allowed, held, denied, redacted, or routed, along with the reasons available at that boundary.
AI attestation
Verifiable evidence for
governed AI decisions.
SASKI connects a deterministic rulebook decision to an inspectable record of what was evaluated, which outcome was recorded, and why.
Plain definition
What SASKI means by AI attestation.
AI attestation is evidence produced at the governance boundary, not an explanation generated afterward by the model.
SASKI separates an AI system’s proposal from the authority to proceed. The rulebook evaluates the relevant interaction or action and records the resulting governance decision.
This creates evidence for engineering, security, risk, and review teams while keeping the limits of that evidence explicit.
Decision record
What an attestation can connect.
- Request reference
- A reference or digest that connects the record to the governed interaction or proposed action without requiring sensitive content in the receipt.
- Rulebook identity
- The policy or rulebook version used when the governance decision was made.
- Decision and reasons
- The allowed, held, denied, redacted, or routed outcome and the reason codes produced by the governed path.
- Authority and timing
- The relevant identity, approval state, and timing information available at the decision boundary.
- Evidence references
- References to related approval or execution evidence when the integrated system provides it.
Two records, two questions
What was decided, and what actually happened?
Execution record
Records what the integrated system reports occurred after the decision. The two records can be linked without treating them as the same claim.
Beyond tool calls
Attestation for actions and high-risk conversations.
Consequential AI does more than invoke tools. It can influence a person before any external action occurs.
SASKI Agentic governs inputs, proposed actions, and outputs. SASKI SDK applies rulebooks around human-facing model interactions. In both cases, the useful evidence is the rulebook decision and its reasons, not a model-generated assurance that it behaved safely.
Explore SASKI Agentic ↗Explore SASKI SDK ↗
Verification boundary
What a signed receipt proves, and what it does not.
A decision receipt can establish
What the governance boundary recorded.
- Which rulebook was identified for evaluation
- Which outcome and reason codes were recorded
- Whether a signed record has changed since signing
- Which related evidence references were retained
It does not establish by itself
Everything that happened afterward.
- That every relevant path passed through governance
- That the original inputs were complete or truthful
- That an allowed action was successfully executed
- That the rulebook was legally or ethically sufficient
Receipt profiles
One engine. The receipt format you need.
SASKI applies the same deterministic rulebook path independently of the receipt profile. The evidence can fit an auditor, insurer, regulator, or existing audit system without changing how the governance decision is made.
Ed25519-signed decision receipts
Each receipt binds the agent and session, requested-action digest, rulebook version and digest, verdict, reason codes, approvals, chain position, build identity, time, and receipt profile. Agentic verifies receipts and exported audit packages offline against pinned public keys.
HMAC attestations with optional Ed25519
Each decision binds product identity, build, time, policy fingerprints, keyed request commitments, decision hashes, chain position, and format metadata. An Ed25519 signature over the same fields can be added for public-key verification. The audit report tool verifies both formats offline.
Signed execution evidence
Estate uses Ed25519-signed execution receipts for governed smart-home actions. Optional external anchoring can sign chain checkpoints with a key held in Google Cloud KMS.
Public test vector
Inspect an Agentic 1.3 decision receipt.
Test-only / Deny decision
Download the signed receipt exactly as issued.
This sanitized Agentic sample includes the requested-action digest, rulebook version and digest, verdict, reason code, build identity, time-window data, receipt profile, key identifier, and Ed25519 signature.
Download receipt-sample.jsonView JSONSample receipts, test vectors, and verifier on GitHub ↗- Product
- SASKI Agentic 1.7.2
- Receipt format
- 1.3 / saski-native@1
- Decision
- Deny / VECTOR_DENY
- Signature
- Ed25519 / vector-1
This vector contains no decimal-number fields. The published test key and verifier reproduce the exact encoding used for this sample.
See the receipt anatomy and verify it yourself ↗A receipt format guide, sample records, test vectors, and a free standalone verifier are public today on GitHub under the Apache 2.0 license. No SASKI account or software is needed to check them.
Integrity and privacy
Continuous evidence without raw content.
Agentic and SDK chain their records. Altering, removing, or reordering a record breaks verification from that point forward.
SDK attestations contain keyed commitments and hashes rather than raw message text. Agentic rejects raw tool arguments at signing. Human-readable display text is added after signing and excluded from audit exports by default.
Records use SASKI canonical JSON: sorted keys, compact separators, UTF-8, and fixed eight-decimal strings for floats. This encoding is intentionally distinct from RFC 8785. Ed25519 signatures use base64; HMAC signatures use hexadecimal encoding.
In child, student, patient, therapist, and HR recruiting modes, the SASKI SDK will not send a request to the AI model unless its decisions are being signed.
Output roadmap
OVERT-aligned receipts are coming soon.
Available today
SASKI native receipts
Agentic, SDK, and Estate produce product-specific attestation. Each record identifies its format, encoding, and signature algorithm; Agentic and SDK include chain and profile metadata inside the signed record.
Coming soon
OVERT-aligned output
SASKI is developing an output profile designed to align with OVERT, the open standard for verifiable runtime evidence in AI governance. The planned profile changes the evidence format, not the deterministic rulebook decision that produces it.
Read the OVERT standard ↗Questions
AI attestation, in practical terms.
What is AI attestation?
AI attestation is durable evidence about a governed AI interaction or action. In SASKI’s architecture, the record connects a rulebook evaluation to its outcome and reasons so that the decision can be inspected later.
What is a signed decision receipt?
A signed decision receipt is a machine-verifiable statement issued at a governance boundary. A valid signature can establish who issued the statement and whether its signed contents changed. It does not, by itself, prove that a later real-world action occurred.
How is a decision receipt different from an execution receipt?
A decision receipt records what governance allowed, held, denied, or routed. Execution evidence records what an external system reports actually happened. Strong accountability connects the two while preserving the distinction.
Does SASKI attestation apply only to agent tool calls?
No. SASKI’s architecture addresses proposed agent actions and human-facing AI interactions. This includes high-risk conversations in which the governed outcome may be to allow, redact, route, hold, or deny a response.
Which signing algorithms does SASKI use?
SASKI Agentic signs decision receipts with Ed25519. SASKI SDK decision attestations use HMAC SHA-256 and can add an Ed25519 signature over the same signed fields for public-key verification.
Can SASKI receipts be verified independently?
Yes. Agentic and SDK ship offline verification tools. Agentic verifies receipts and audit packages against pinned public keys. The SDK audit report tool verifies chains, HMAC attestations, and Ed25519 signatures. Neither requires a network connection or SASKI account. The sample receipts, test vectors, and verifier are public on GitHub.
Start with your use case