@bedrockgovernance/attest

Record what an AI model was asked and what it produced, at the moment of generation, as an immutable entry in the Bedrock advice ledger.

@bedrockgovernance/attest captures a generation, the event of a model producing text that helps shape a piece of financial advice, and notarises it into the firm's ledger. Where CRM integrations ingest the finished advice document after it exists, attestation captures the model call at its source: the system prompt, the conversation, the retrieved context, the guardrails, and the output, hashed and chained at the instant they are used.

The client has no runtime dependencies. The optional AI SDK middleware lives in a separate entry point (@bedrockgovernance/attest/ai-sdk), so importing the core never pulls in the AI SDK.

Install

npm install @bedrockgovernance/attest

Quickstart

Attest explicitly

Hand Bedrock the generation bundle yourself:

import { Bedrock } from '@bedrockgovernance/attest';

const bedrock = new Bedrock({ apiKey: process.env.BEDROCK_API_KEY! });

const { generationId } = await bedrock.attest({
  // A stable id you already hold that scopes this drafting
  // conversation, e.g. a chat or thread id. Reuse it on every
  // generation in the same conversation so they group together.
  correlationId: chatId,
  clientReference: 'CLI-41269355',
  adviser: { name: 'Jane Smith', fcaRef: 'JXS01234' },
  model: { provider: 'anthropic', name: 'claude-opus-4-8' },
  instructions,
  input,
  retrievedContext,
  output: { content: draft, finishReason: 'stop' },
});

Only correlationId, model, input and output are required. input is what the model saw for this call: a single prompt string, or the role-tagged messages sent. instructions (the system prompt) is optional but worth including whenever your call has one. Everything else is optional too, included when your pipeline has it.

Attest automatically with the AI SDK

Wrap your model once and every generateText / streamText call is notarised, with no change to your generation code:

import { wrapLanguageModel } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';
import { attestMiddleware } from '@bedrockgovernance/attest/ai-sdk';

const model = wrapLanguageModel({
  model: anthropic('claude-opus-4-8'),
  middleware: attestMiddleware({ bedrock, correlationId: chatId, adviser }),
});

The middleware posts the attestation in the background after the model responds, so it never adds latency to the call. Attestation failures are surfaced through an optional onError callback and never affect the model result. When onError is omitted the middleware will log a console warning by default so background attestation failures are discoverable unless you handle them explicitly.

Grouping generations

A single piece of advice is rarely one model call, and the calls are often exploratory chat turns rather than drafts of the final document. correlationId is how you group them: pass a stable id you already own (a chat, thread, or case id) on every attest() call that belongs to the same conversation. It does not need to be globally unique, only stable across the conversation. Bedrock stitches conversations into a piece of advice on the platform side, so the id does not have to survive to review time.

API

• new Bedrock({ apiKey, baseUrl?, fetch? }) — construct a client. The apiKey is sent as the x-bedrock-key header. baseUrl defaults to the production API; fetch defaults to the global fetch (Node 18+).
• bedrock.attest(generation) — validate the bundle and record it via POST /v1/generations. Resolves to { generationId, correlationId, outputHash, recordedAt }.
• attestMiddleware(options) (from @bedrockgovernance/attest/ai-sdk) — an AI SDK LanguageModelV2Middleware that attests every call.
• validateGeneration(generation) — the same local validation attest runs, exported for reuse.

Errors: BedrockValidationError (invalid bundle or config, thrown before any request), BedrockApiError (non-2xx response, carrying status, code and requestId), and their base BedrockError.

License

Apache 2.0. See SECURITY.md for vulnerability reporting.