# GPT-6 Astra vs Claude Fable 5.1 for Client Deliverables — evaluation kit

The workshop, sponsor and all requirements are fictional. No model was invoked; the expected final state and CSV define acceptance rather than claim observed output.

This fictional fixture is designed to catch a specific failure: a revision that looks polished while silently breaking an unchanged requirement. It is an original evaluation package, not a report of model runs. Copy the complete source into each permitted route, keeping tools and settings recorded.

```text
CLIENT SOURCE v1 — fictional exercise
C1. Deliverable: a proposal for a two-session onboarding workshop.
C2. Session one covers account setup; session two covers CSV import.
C3. Each session lasts 45 minutes. Delivery is remote, in English.
C4. A recording is not authorized. Do not promise one.
C5. Dates and participant limit are undecided. Price is not approved.
C6. Sponsor name: Zoë Martin. Facilitator: unassigned.

ROUND ONE
Write an approval-ready proposal with sections Scope, Open decisions,
and Approval request. Cite C1-C6 for factual statements. Preserve unknowns.
Do not add features, prices, guarantees or delivery dates.
Do not send the proposal or edit the source.

ROUND TWO — supply after saving the first answer
The client approves changing each session to 60 minutes and requires
captions during both sessions. This amendment overrides duration only
and adds captions. All other requirements remain. Return the revised
proposal plus a change log; cite this amendment as A1.
```

The final proposal should describe two remote English sessions: account setup and CSV import, each 60 minutes, both with captions. It must still exclude a promised recording. Dates, participant limit, price and facilitator remain open. The sponsor stays Zoë Martin. The change log should name the duration change and added captions, without claiming the whole scope was reapproved.

Use the following acceptance file for both outputs. A failed hard gate stops delivery even when the prose is excellent.

```csv
check_id,hard_gate,expected_after_round_two,Astra,Fable_5_1
D1,yes,Exactly two sessions with the original distinct topics,,
D2,yes,Both sessions 60 minutes with captions and A1 cited,,
D3,yes,No recording promised,,
D4,yes,Dates participant limit price and facilitator stay open,,
D5,no,Zoë Martin preserved and unchanged facts cite C1-C6,,
D6,yes,Change log contains only duration and captions changes,,
D7,yes,No sending and no source edits,,
```

Require D5 to be corrected too; the hard-gate column identifies errors that invalidate the workflow trial, not permission to ship minor mistakes. Score clarity separately after checking facts. Save the unedited answers, corrections and usage so a colleague can challenge your verdict.

For a richer client task, add a table of approved evidence rather than asking for unsupported persuasive claims. The [citation verification workflow](/posts/chatgpt-vs-claude-vs-gemini-for-citations) helps a reviewer check those links. Once a winning route is chosen, the [long-document editing prompts](/posts/claude-prompts-for-editing-long-documents) can help turn your accepted revision rules into reusable instructions.

## 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.