Pillar 02 — Verifiable AI

Verifiable AI.

Independently reproduce and verify AI decisions, after the fact.

Record AI inputs and outputs, the model ID, and the applied policy as cryptographic proofs. Leave the decision process in a form a third party can reproduce without ever accessing the original. One of the four pillars that make up Lemma's trust infrastructure.

Answering the questions being asked most urgently today

Held accountable for AI decisions under regulation and audit

Need to guarantee decision consistency across model switches

Need to detect tampering with AI inputs after the fact

01 · What is Verifiable AI

Make "what it saw and how it decided" independently verifiable, after the fact.

Issue commitments to the AI's input, the model used, and the output. A third party can verify the reproduction and compliance of the decision without accessing the original.

AI workflow

Input + model + output

The input data the AI processes in your workflow, the model used, and the output result.

The original data stays in-house

Lemma

Commitment + policy verification

1Commit the input hash

2Record the model ID and policy

3Guarantee consistency with a ZK proof

Never touches confidential inputs or model internals

Audit / regulatory response

A verifiable decision log

modelIdgpt-4o

inputCommitment0xa3...

outputCommitment0xb7...

satisfiesPolicy✓

A third party reproduces and verifies it independently

Think of it like meeting minutes.

After an important meeting, you keep minutes recording who attended, what was discussed, and what was decided — sound familiar? Lemma's Verifiable AI is a mechanism that automatically issues these minutes for every AI decision. Later, a third party can verify that "this input really did produce this decision from this model," without ever seeing the original.

What you disclose

A verifiable fingerprint of the decision

✓modelId (identifier of the model used)
✓inputCommitment / outputCommitment
✓satisfiesPolicy (policy compliance)
✓A ZK proof for third-party verification

What this mechanism makes possible

✓AI decisions are kept as reproducible "minutes"

✓The input itself and model internals stay protected as confidential

✓You can meet accountability requirements in regulation and audit

✓You can verify decision consistency even across model switches

Meet AI accountability with cryptography.

→Build workflows where regulatory audit holds up retroactively

→Keep decisions consistent even across model switches

→Automation and verifiability coexist by the same construction

02 · How it differs from existing approaches

How it differs from AI logging, model audit, and access control.

Merely "recording," "auditing," or "controlling" does not guarantee tamper-proofness and independent verification.

Approach

Reproduce the decision

Confidential data stays local

Independent verification

Cryptographic proof

AI logging only

△

✓

✗

Model audit tools

✗

✓

△

✗

Access control only

✗

✓

✗

Lemma Verifiable AI

✓

03 · Position in the trust infrastructure

P2 is one of the four pillars that make up Lemma's trust infrastructure.

P1 Origin P2 AI P3 Authority P4 Attributes

See the full trust infrastructure →

04 · In practice

Workflow scenarios that use this pillar.

P2 · Log

AI audit log proof

P2 · Attack

Prompt injection detection

P2 · Regulation

EU AI Act compliance

P2 · Model

Model integrity proof

P2 · Document

AI document isolation

P2 · RAG

RAG source attestation

See all use cases →

05 · Technical reference

Make AI decisions verifiable with proprietary cryptography.

Commitment

Fixes the contents of inputs and outputs as a unique fingerprint without revealing them. You can independently verify later whether "it was the same input."

Policy verification

Proves with ZK whether the AI's decision satisfies a predefined policy (age limits, geo restrictions, and so on).

Model ID fixing

Cryptographically fixes which model (version, parameter set) performed the inference. For tracking consistency across model switches.

// your.model.v1 — Lemma canonical schema
{
  "$schema": "your.model.v1",
  "modelId": "gpt-4o-2024-08-06",
  "inputCommitment": "0xa3f1...c9d2",
  "outputCommitment": "0xb7e2...4a8f",
  "policyHash": "0xf1c5...e3b9",
  "satisfiesPolicy": true,
  "timestamp": "2026-06-05T10:00Z",
  "zk_proof": "0x8c4f...e7d2"
}

Developer guides for implementing P2