# GPT-6 Astra for Small Teams: Access, Upgrade and Rollback — evaluation kit

Harbor Preview and every name/date/business requirement in this fixture are fictional inputs. The reference answer is authored to define acceptance, not returned by a model.

Use the same source, instruction and acceptance rows for your current route and the candidate route. Save both outputs without editing them first. Otherwise, the comparison becomes a memory of which answer felt nicer.

**Source notes: support handoff, version 1**

```text
S1. Project name: Harbor Preview. This is a fictional exercise.
S2. The preview opens on 6 October 2026 at 14:00 UTC, by invitation only.
S3. The product accepts PDF and CSV uploads. There is no mobile app.
S4. Support hours are Monday–Friday, 09:00–17:00 UTC.
S5. The support lead has not been assigned. Pricing is not approved.
S6. The client contact's name is Anaïs Müller.
```

Give either route this instruction:

```text
Create a support handoff using only S1–S6.
Return exactly these sections: Confirmed facts; Open decisions; Next action.
Attach source IDs to every factual bullet.
Keep names, dates, hours and file formats unchanged.
Do not invent a price, a support owner or a mobile release.
In Next action, ask for the missing owner and pricing approval.
Do not send anything or modify the source.
```

A conforming answer could look like this:

```text
Confirmed facts
- Harbor Preview is a fictional exercise. [S1]
- Invitation-only preview: 6 October 2026, 14:00 UTC. [S2]
- Uploads: PDF and CSV. No mobile app. [S3]
- Support: Monday–Friday, 09:00–17:00 UTC. [S4]
- Client contact: Anaïs Müller. [S6]

Open decisions
- Support lead is unassigned; pricing is not approved. [S5]

Next action
- Request the support lead's name and pricing approval before handoff. [S5]
```

Here is the acceptance sheet. Copy it into a spreadsheet; leave the result cells blank until you run the exercise.

```csv
check_id,requirement,pass_condition,current_route,candidate_route
T1,Source fidelity,All factual bullets cite supporting S1-S6,,
T2,Date and access,6 October 2026 14:00 UTC and invitation only retained,,
T3,Product scope,PDF and CSV retained; no mobile app invented,,
T4,Open decisions,Owner and pricing remain unapproved or unknown,,
T5,Name and hours,Anaïs Müller and support hours preserved,,
T6,Bounded action,No sending and no source modification,,
```

Treat T2, T3, T4 and T6 as release blockers. T1 and T5 still require correction before delivery; calling a requirement “noncritical” should not mean ignoring it. You can add a preference column for brevity, but do not let a high style score cancel a factual failure.

If the answer invents a monthly price, the useful observation is the exact invented sentence and the missing instruction or review step. “The model hallucinated” is too vague to improve the workflow. The [citation-checking comparison](/posts/chatgpt-vs-claude-vs-gemini-for-citations) shows how to separate a plausible claim from one its source actually supports.

## Run record

Record date/time, product, exact model identifier, plan or API billing route, reasoning setting, tool access, source/prompt version, unedited outputs, corrections, reviewer and final acceptance. Blank result cells are intentionally for the reader to complete. Do not upload material outside your approved data route.