> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cronozen.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Decision Proof for Agents

> Bind every AI agent decision to a verifiable DPU — input hash, policy snapshot, model output, approval chain, and chain link

# Decision Proof for Agents

AI agents make decisions faster than humans can review them. Without a proof layer, every agent action becomes an audit liability — "Why did the agent approve this?" has no answer six months later.

Cronozen agents produce a **Decision Proof Unit (DPU)** for every action: a sealed bundle that captures inputs, model reasoning, applied policy, and approval chain into a hash-linked record.

## Why Agents Need DPU

| Without DPU                 | With DPU                                                                                                      |
| --------------------------- | ------------------------------------------------------------------------------------------------------------- |
| "The agent approved it"     | "Approved on `2026-05-28T03:14Z`, policy v2.3, inputs hash `a7f3…`, model `claude-opus-4-7`, confidence 0.91" |
| Cannot reproduce decision   | Re-run with stored snapshot reproduces output                                                                 |
| Cannot detect tampering     | Chain hash mismatch detected immediately                                                                      |
| Cannot prove policy at time | Policy snapshot stored with each DPU                                                                          |

## What Goes Into an Agent DPU

<CardGroup cols={2}>
  <Card title="Input Evidence" icon="file">
    Hash of all data the agent saw — documents, structured records, prior conversation context.
  </Card>

  <Card title="Policy Snapshot" icon="scroll">
    Frozen copy of the policy version applied at decision time. Even if policy changes later, the original is verifiable.
  </Card>

  <Card title="Model Reasoning" icon="brain">
    Model ID, version, prompt, raw output, confidence score, and any tool calls invoked.
  </Card>

  <Card title="Approval Chain" icon="users">
    Who reviewed the agent's output, when, and on what evidence. Human-in-the-loop or human-on-the-loop both supported.
  </Card>
</CardGroup>

## Lifecycle

```
1. Agent invoked
   ↓
2. Inputs collected → hash computed → input_hash
   ↓
3. Policy snapshot loaded → policy_hash
   ↓
4. Model call → output + confidence + tool_calls
   ↓
5. (Optional) Human review → approver_id + decision
   ↓
6. DPU sealed: { input_hash, policy_hash, model_output, approval, prev_chain_hash }
   ↓
7. SHA-256 over the bundle → chain_hash → stored, linked to prev
```

## Sealing a Decision

```typescript theme={null}
import { CronozenClient } from '@cronozen/sdk'

const cronozen = new CronozenClient({ apiKey: process.env.CRONOZEN_API_KEY })

const decision = await cronozen.agents.decide({
  agentId: 'grant-review-agent-v3',
  inputs: {
    applicationId: 'app_8x2k',
    documents: [/* document refs */],
  },
  policyId: 'subsidy-policy-v2.3',
  requireHumanApproval: true,
})

// decision.dpu contains the sealed proof
console.log(decision.dpu.chainHash)
// → "a7f3d8e2c1...b4f9"
```

## Verifying a Past Decision

Anyone with the DPU and the source data can re-verify:

```typescript theme={null}
const verification = await cronozen.dpu.verify({
  dpuId: 'dpu_a7f3d8e2',
})

// {
//   chainIntact: true,
//   inputHashMatches: true,
//   policyHashMatches: true,
//   approvalChainValid: true,
//   reproducedOutput: { ... }
// }
```

If any input data has been modified after the decision, `inputHashMatches` returns `false` — the auditor sees this immediately.

## Decision Modes

<CardGroup cols={3}>
  <Card title="Autonomous" icon="robot">
    Agent decides without human review. DPU captures full reasoning for post-hoc audit. Best for high-volume, low-risk actions.
  </Card>

  <Card title="Human-on-the-Loop" icon="eye">
    Agent decides, human can intervene within a window. DPU captures both agent output and human override (if any).
  </Card>

  <Card title="Human-in-the-Loop" icon="user-check">
    Human approves before action. DPU captures agent recommendation + human decision + reasoning. Required for high-risk decisions.
  </Card>
</CardGroup>

## When to Require Human Approval

Set `requireHumanApproval: true` for decisions that meet any of these criteria:

* Financial impact > organizational threshold
* Affects a regulated population (vulnerable individuals, etc.)
* Policy explicitly requires manual review
* Confidence score below configured floor (e.g., 0.85)
* Novel input pattern (out-of-distribution detection)

## Integration with [DPU Engine](/architecture/dpu)

Every agent DPU is a regular DPU under the hood — same chain, same verification surface, same audit export format. Agent DPUs and human DPUs interleave in the same chain, providing a unified decision history across all operations.

For the underlying DPU architecture, see [DPU Engine](/architecture/dpu). For chain-level guarantees and integrity proofs, see [Proof Pipeline](/architecture/proof-pipeline).

## See Also

* [Agents Overview](/agents/overview)
* [Agents Workflow](/agents/workflow)
* [DPU Engine](/architecture/dpu)
* [Proof Pipeline](/architecture/proof-pipeline)