---
title: "Verifiable by Design"
description: "Source attribution, audit trails, inspectable reasoning — what makes output trustable."
order: 18
duration: "5 min"
chapter: "05-trusting-your-own-data"
type: lesson
---

## What makes AI output trustable

In Cowork, you can scroll up and see exactly what code Claude ran on your data. That is not a minor UI feature — it is trust infrastructure. Three things make AI output trustable:

**Source attribution.** Where did this data come from? When Claude says "Chen's Produce charged sales tax on three February invoices," you need to see which transactions it is referencing. The data has to be traceable to the source.

**Audit trails.** What steps did the system take? If Claude ran a script to aggregate transactions by vendor and month, you need to see that script. Not because you distrust Claude, but because you need to verify the methodology, not just the result.

**Inspectable reasoning.** Why did it reach this conclusion? When Claude flags a duplicate charge, it should show the two transaction records, the matching amounts, the close dates — the evidence, not just the verdict.

This infrastructure does not exist in a chat window. When you ask Claude.ai "are there any errors in my transactions?" and paste a few rows, you get an answer but no trail. Cowork is the step up — it runs code you can inspect, on data you can verify, with intermediate results you can examine.

MCP takes this further. It does not just connect your data — it makes the connection inspectable. Every tool call is logged. Every data retrieval has a source. Every action can be traced back to the event that triggered it.

## Check your understanding

Answer in your own words — write it down before opening the key. Your tutor grades against the criteria and generates fresh variants on retries.

**Q1.** Name the three ingredients of trustable AI output, and for one of them give a concrete example of what it looks like when Claude analyzes a spreadsheet of your data.

<details>
<summary>Answer key — attempt every question first</summary>

## Answer key

### Q1

**Model answer:** Source attribution (which rows produced this claim), audit trails (the actual script/steps run), inspectable reasoning (the evidence behind each flag). Example for audit trails: in Cowork you scroll up and read the exact code Claude ran to aggregate transactions — you verify the methodology, not just the verdict.

**Pass criteria:** all three ingredients named; example ties one ingredient to something observable

</details>


**Next:** [Connectors: One Click to Your Data](/mastering-claude/05-trusting-your-own-data/19-connectors/)