Fix eval set for GPT 53 upgrade

[harsheetjain]

Why prompt changes are needed between model upgrades

Eval prompts are not model-agnostic. Each model generation interprets intent,
phrasing, and ambiguity differently — a query that resolves cleanly on one model can
become ambiguous on the next.

In this case, GPT-5.3 was reading the yes/no framing ("Can I…?") as a
permissions question rather than the intended info + update-process intent.
The reworded query makes both sub-intents explicit.

🔑 When we upgrade the underlying model, eval sets must be re-tuned so the
scorecard measures agent quality, not phrasing artifacts the new model handles
differently. Otherwise regressions reflect prompt drift, not real behavior changes.

---

Test plan

• Re-run ESS eval suite against GPT-5.3
• Confirm the reworded query scores ≥ threshold on CompareMeaning
• Spot-check no regression on prior model baseline

[harsheetjain]

@microsoft-github-policy-service agree company="Microsoft"

[nkemms]

👍