THE CUSTOMER01
“Your AI promised us a refund.”
A screenshot appears. Your logs show a different answer. Neither side can prove which record changed.
A mutable log cannot settle this.
AI / 01SIGNED MODEL EVIDENCE
When an AI answer matters, a log is not enough.
The dispute will not be about what your model usually does. It will be about one exact prompt, one exact answer, and one moment your database alone cannot prove.
Invoance turns that moment into signed evidence anyone can verify independently.
50 attestations / month freeOne call after generationNo proxy in your model path
DISPUTE RECORDclaim_48391 / decision
VERIFIED
QUESTION UNDER REVIEW
Should this commercial claim be escalated?
MODEL RESPONSE
Escalate. Confidence 0.93. Policy threshold 0.81.
MODELgpt-4.1
ANCHORED14:32:07.123Z
PROOF IDatt_01JQ7F…P8E
INPUT
7db0a6f43ef98c21…OUTPUT
21be74f89a67c011…ED25519 SIGNATURE
VALID6fB3kYw9n2M4dQ7t…eA91The reviewer checks the math. They do not have to trust your dashboard.
EXACT INPUTEXACT OUTPUTMODEL + VERSIONSIGNED RECORDINDEPENDENT VERIFICATION
02THE DAY SOMEONE ASKS
The hard question arrives after the context is gone.
Your logs were built for debugging. The person challenging an AI outcome needs evidence that survived your own ability to edit it.
THE REVIEWER02
“Why did the model deny this claim?”
The model version moved on, the prompt template changed, and the exported row depends on your word.
A mutable log cannot settle this.
THE INCIDENT03
“What exactly did the agent do?”
Ten systems hold fragments of the run. The moment that matters is buried between mutable events.
A mutable log cannot settle this.
03THE DIFFERENCE
Turn “we think this happened” into “verify it yourself.”
APPLICATION LOGA record you control
14:32:07
decision = "escalate"14:35:19
record updated—
Who changed it?- Editable by your systems
- Origin depends on your word
- Export still requires trust
SIGNED ATTESTATIONA record the math controls
INPUT HASH
7db0…98c2OUTPUT HASH
21be…c011SIGNATURE
6fB3…eA91- Any signed change is detectable
- Organization and moment are bound
- Third parties verify independently
04FROM ANSWER TO EVIDENCE
Your inference path stays untouched.
Send a copy after the model answers. Your customer gets the response first; attestation happens beside the product, not in front of it.
01
MODEL RESPONDSYour normal AI workflow completes02
ONE ASYNC CALLInput, output, and model context are sent03
PROOF RETURNSHashes and a durable attestation idNo model proxy. No inference dependency.One API call after generation.
05SHIP IT TODAY
One call. Your language. A proof back.
Official SDKs keep the integration small. Start with the exchanges most likely to be questioned, then expand when the evidence earns its place.
SELECT SDK
Read the implementation guideNode.js / CREATE ATTESTATIONPOST-GENERATION
REQUEST
const proof = await client.attestations.ingest({
type: "decision",
input: prompt,
output: response,
modelProvider: "openai",
modelName: "gpt-4.1",
modelVersion: "2026-04-14"
});201 RESPONSE
{
attestation_id: "att_01JQ7F...P8E",
input_hash: "7db0a6f43ef9...",
output_hash: "21be74f89a67...",
payload_hash: "f93c0d51c240...",
status: "accepted"
}06WHAT CHANGES
Evidence changes the conversation.
The value appears when an important AI outcome is questioned. Instead of rebuilding history, you present a record.
01
Resolve disputes
Show the signed exchange instead of arguing over screenshots, exports, or reconstructed timelines.
02
Answer assurance reviews
Give reviewers a portable record they can verify without dashboard access or blind trust in your database.
03
Trace agent decisions
Attach attestations to a trace and preserve the model calls that shaped the final action.
Exact input
SHA-256 hash of what the model received
Exact output
A separate hash of what the model returned
Model identity
Provider, model name, and version in the signed payload
Moment
An anchored timestamp bound to the record
Origin
Ed25519 signature scoped to your organization
Independent check
Portable verification material for third parties
07BEFORE YOU SHIP
The questions buyers ask.
01Does this slow down the model response?
No. Attestation is designed to run after generation on a copy of the exchange, so it does not need to sit in the inference path.
02Is this just another application log?
No. The input, output, and canonical payload are hashed, and the record is signed with an organization-scoped Ed25519 key so later changes are detectable.
03Can someone verify without our dashboard?
Yes. Verification material is portable, and supported SDK flows can check signatures independently rather than trusting a dashboard display.
04Does attestation prove the model was correct?
No. It proves what was presented, what the model returned, which model context was recorded, and whether the signed record changed. It does not certify factual accuracy or legal compliance.
READYPROTECT THE NEXT ANSWER
The cheapest proof is the one you create before the dispute.
Start with the AI decisions your customers, reviewers, or team are most likely to question.
50 attestations per month free · No credit card required