# AI Execution V2 through Evidence Bridge

URL: https://docs.nexart.io/docs/evidence-bridge/ai-execution

Map an existing model-execution payload to the cer.ai.execution.v2 evidence family: required semantics, conceptual field mapping, reviewed target configuration and confidential sealing.

Use the AI Execution target when the source represents an actual AI or model execution. The canonical record family is `cer.ai.execution.v2` at protocol `1.3.1`, produced by `@nexart/ai-execution`. See AI Execution V2 for the normative schema.

## Required semantics

- execution identity
- execution timestamp
- provider
- model
- input
- output
- parameters
- execution role

Optional fields may include instruction, instruction kind, model evidence, decision references and extensions.

## Source example

Source payload

```
{
  "run_id": "run_9281",
  "timestamp": "2026-09-09T10:20:30.000Z",
  "provider": "openai",
  "model": "gpt-4.1",
  "prompt": {
    "customer": "Example request"
  },
  "response": {
    "classification": "approved"
  },
  "temperature": 0
}
```

## Conceptual mapping

```
run_id       -> executionId
timestamp    -> timestamp
provider     -> provider
model        -> model
prompt       -> input
response     -> output
temperature  -> parameters.temperature
```

Field names alone are not sufficient. Evidence Bridge requires a reviewed target configuration so that arbitrary JSON fields cannot automatically acquire AI execution semantics merely because they are named "input", "output" or "model".

## AI confidentiality

Production AI evidence is confidentially sealed before Node attestation. Protected plaintext such as model input and output is replaced by cryptographic commitment envelopes in the public CER:

Public representation of a protected field

```
{
  "_redacted": true,
  "mode": "confidential",
  "commitmentScheme": "hmac-sha256-v1",
  "commitment": "hmac-sha256:..."
}
```

Private opening material is kept separately. The Node receives the confidential CER, not the private openings. An opening can later be used for selective verification of the committed value.

NexArt confidentiality uses keyed, salted HMAC commitments managed by the canonical SDK. A plain SHA-256 hash does not provide confidentiality.

See Confidential evidence for openings and selective verification.