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

# Module 5 · The Same-Question Test

> Goal: the proof moments — the same broad question under two personas, the off-scope decline traced to a readable file, and a behavior change made through governance. (~20 minutes)

**Goal:** the proof moments — the same broad question under two personas, the off-scope decline traced to a readable file, and a behavior change made through governance.

<Note>
  Setup: your two Module 4 personas approved and provisioned to two test identities (or two sessions an admin can switch — Module 6 covers assignment mechanics). Have both sessions open side by side.
</Note>

## 5.1 · The side-by-side

Ask the **identical broad question** in both sessions:

```text Prompt theme={null}
How are our [members/patients] doing on [hospital utilization] and cost?
```

<Check>
  **You'll see:** Ana recognizes the breadth and *clarifies with a menu of allowed options only* ("through care gaps, utilization, or outreach — which would help most?"). After you pick: plain language, summary first, **no SQL anywhere**, sources described in business terms.
</Check>

<Check>
  **You'll see:** Ana *proceeds with stated assumptions* ("assuming utilization means admissions + ED visits over the trailing 12 months..."), answers in full technical detail, **shows the SQL** with a clause-by-clause explanation, and flags anything computed beyond the governed surfaces.
</Check>

Same organization, same ontology, same question — two correct experiences. Capture both screens; this side-by-side is the single best artifact for explaining the context stack to anyone who wasn't in the room.

## 5.2 · The off-scope decline — and the file behind it

In the **restricted** session, ask for something real but outside the allowlist:

```text Prompt theme={null}
Show me the [unit-cost breakdown by provider contract / financial detail] for last quarter.
```

<Check>
  **You'll see:** a warm decline with a redirect — "that falls outside this workspace; contact \[the analytics team]. Within this workspace I can help with..." — and *no* hint of what other surfaces exist.
</Check>

Now the trust moment. Open `behaviors/[program_manager]/org_context.md` and read section 4 aloud next to the decline you just received. **The behavior you witnessed is a sentence in a version-controlled file** — not a vibe, not a model mood. It's reviewable, diffable, auditable: when compliance asks "what stops a program manager from seeing financials?", the answer is a file, its review history, and the enforcement layers beneath it.

## 5.3 · Change behavior through governance

The program-management team now legitimately needs \[readmission detail]. Wrong move: someone edits the file directly and behavior drifts. Right move:

* **Propose the patch:** add `[readmissions_detail]` to the persona's allowed-surfaces list — one line — with a description citing who requested it and why.- **Review and approve** (the metric's or domain's owner signs off, per your Module 3 review bar).- **Re-ask** in the restricted session — the previously declined question about readmission detail now answers, in the persona's plain-language style.

<Check>
  **You'll see:** behavior change with a paper trail — *who* requested it, *who* approved it, *when*, and exactly *what* changed. Compare that against the alternative ("an admin toggled something in a console six months ago, nobody remembers why") and you have the whole argument for context-as-files.
</Check>

## 5.4 · Write the test down

The same-question test isn't just a demo — it's your regression suite for personas. Record the three probes (the broad question, the off-scope request, the newly-allowed request) and their expected behaviors; re-run them whenever a persona file changes. Module 7 makes this a quarterly habit.

<Frame caption="Same question, two personas — scope and voice differ, the governed truth does not.">
  <img src="https://textqllabs.github.io/workshops/context-stack/assets/ctx-m4-sidebyside.png" alt="Same question, two personas — scope and voice differ, the governed truth does not." />
</Frame>

<Frame caption="The off-scope decline — access control you can read.">
  <img src="https://textqllabs.github.io/workshops/context-stack/assets/ctx-m4-decline.png" alt="The off-scope decline — access control you can read." />
</Frame>

### ✅ Checkpoint

* [ ] The same broad question produced clarify-vs-proceed, no-SQL-vs-SQL, summary-vs-assumptions — captured side by side
* [ ] The off-scope decline matched section 4 of the persona file, word for behavior
* [ ] The allowlist change went through propose → review → approve, and the re-ask confirmed it
* [ ] The three-probe test is written down as the persona's regression suite