Inference Receipts

The @peac/adapter-openai-compatible package creates hash-first interaction evidence from chat completion API responses. It records SHA-256 digests of prompts and outputs -- no raw text is ever stored in the receipt.

Privacy by default

Inference receipts use SHA-256 hashes exclusively. The receipt proves that a specific interaction occurred without revealing the content of prompts or completions.

---

Install

Terminal

pnpm add @peac/adapter-openai-compatible

---

Quick start

inference-evidence.ts

import { fromChatCompletion } from '@peac/adapter-openai-compatible';

const response = await openai.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Explain PEAC receipts' }],
});

const evidence = fromChatCompletion(response, {
  messages: [{ role: 'user', content: 'Explain PEAC receipts' }],
});

The evidence object contains:

• Model ID -- which model produced the output
• Token counts -- prompt and completion token usage
• Timing -- response timestamps
• Input digest -- sha256(canonicalized_messages)
• Output digest -- sha256(completion_content)
• Finish reason -- stop, length, tool_calls, etc.

---

What is recorded

Field	Source	Privacy
model	Response model field	Visible
usage.prompt_tokens	Token count	Visible
usage.completion_tokens	Token count	Visible
finish_reason	Completion choice	Visible
input_digest	sha256(messages)	Hash only
output_digest	sha256(completion)	Hash only
created	Response timestamp	Visible

What is NOT recorded

• Raw prompt text
• Raw completion text
• System instructions
• Function/tool definitions
• Image or audio content

---

Hashing functions

For advanced use cases, the individual hashing functions are exported:

hashing.ts

import { hashMessages, hashOutput } from '@peac/adapter-openai-compatible';

// Hash the input messages
const inputDigest = hashMessages(messages);

// Hash the completion output
const outputDigest = hashOutput(completion.choices[0].message.content);

Both functions use SHA-256 with deterministic canonicalization to ensure the same content always produces the same digest.

---

Provider compatibility

The adapter works with any provider that returns OpenAI-compatible chat completion responses:

• OpenAI (GPT-4o, GPT-4, etc.)
• Anthropic (via OpenAI-compatible endpoint)
• Ollama (local models)
• vLLM (self-hosted)
• Together AI
• Any provider with OpenAI-compatible API format

Category-first

The adapter operates on the response format, not the provider. Any API that returns a standard chat completion response object will work.

---

Attaching to a receipt

issue-inference-receipt.ts

import { issueReceipt } from '@peac/protocol';
import { fromChatCompletion } from '@peac/adapter-openai-compatible';

const evidence = fromChatCompletion(response, { messages });

const receipt = await issueReceipt({
  privateKey: process.env.PEAC_PRIVATE_KEY,
  kid: 'peac-2026-02',
  claims: {
    iss: 'https://api.example.com',
    sub: 'agent:assistant-1',
    peac: {
      type: 'inference',
      attestation_type: 'interaction',
      status: 'executed',
      extensions: {
        'org.peacprotocol/inference@0.1': evidence,
      },
    },
  },
});

---

Next steps

Content Signals

Observe content use policy signals before making inference requests.

Evidence Carrier

How inference evidence travels across MCP, A2A, and HTTP transports.

MCP Server

Verify and inspect inference receipts via MCP tools.