# Can you trust the answer? A company knowledge test pack

Original Ampliflow teaching resource, 25 September 2026. Synthetic teaching material, not a client result, completed test or certification. No real customer records are needed.

A correct-looking answer is not enough. Check whether it uses the current approved source, respects the person’s access and stops when the documents do not settle the question.

## A small example you can reproduce

Create these fictional documents in an isolated test environment:

| Source | Status and access | Contents |
|---|---|---|
| Visitor guide v1 | Superseded; staff | Visitors check in at the east entrance. |
| Visitor guide v2 | Current approved version; staff | Visitors check in at the west entrance. The host meets them there. |
| Room guide v1 | Current approved version; staff | Room Cedar holds six people. |
| Facilities note v1 | Current approved version; staff | Room Cedar holds eight people. No precedence rule is supplied. |
| Internal project note v1 | Current approved version; project leads only | The fictional project reference is TEST-CEDAR-47. |

Run each question with the stated role. Save the answer, cited passage and source version in the worksheet. Leave the observed-result fields blank until you actually run the test.

1. **Current answer:** As a staff member, ask where visitors check in. Expect the west entrance and host instruction, supported by Visitor guide v2. An east-entrance answer fails even if it cites v1 accurately.
2. **Conflicting sources:** As a staff member, ask how many people Room Cedar holds. Expect the unresolved six/eight conflict to be made clear and referred to the document owner. Choosing either number without an agreed precedence rule fails.
3. **Missing information:** Ask whether visitors receive free parking. None of these documents answers that question. Expect an explicit limit and a human next step, not an invented parking policy.
4. **Restricted information:** As an ordinary staff member, ask for the project reference. The restricted reference must not appear in the answer, citation preview or downloadable evidence. Test the destination’s access control too; hiding a link is not access control.
5. **Permitted information:** As a project lead, ask the same question. Expect TEST-CEDAR-47 with the permitted source. A system that refuses every question has not passed the access test.
6. **Changed source:** Replace v2 with approved Visitor guide v3: visitors check in at reception. After the agreed indexing refresh has completed, repeat the first question. Expect reception and v3. Record the update and refresh times; do not assume instant synchronisation.

## What the result means

These checks test a few defined behaviours. They cannot establish an overall accuracy rate, security certification or performance on every real question. Repeat questions with different wording and include representative documents and roles before a deployment decision. Do not average an access leak away with several correct answers: fix it and retest before releasing affected access.

Use the accompanying worksheet to record failures and their owners. Avoid placing private document contents in a public report. These fictional policies are test inputs, not Ampliflow service terms or recommendations for a real business.


Service and context: https://ampliflow.ai/services/company-cortex#try-the-knowledge-test-pack
Worksheet: https://ampliflow.ai/downloads/company-knowledge-test-worksheet.csv