Worked benchmark·Demo-safe data·July 2026

Support triage prompt regression: v16 vs v17

A small wording change removed an escalation rule. The candidate looked more specific, but its pass rate fell from 94% to 72% on the same 50 cases.

Baseline

94%

Prompt v16

Candidate

72%

Prompt v17

Pass → fail

11

of 50 cases

Decision

Block

Restore and rerun

The report

The aggregate drop and the evidence behind it

The result is useful because the conditions stayed fixed: GPT-5.4, temperature 0.2, the same 50-case dataset, and the same scoring contract. The prompt diff is the meaningful changed variable.

Support triage release check

9472−22 points

Baseline v16 vs candidate v17 · 50 cases · same model, same dataset

Regression detected

11 cases passed the baseline but fail the candidate.

Prompt changes · v16 → v17

[system]
You are a support triage assistant.
Categorize each ticket as billing,
technical, urgent, or general.
Escalate outages and revenue loss
as "urgent", even when technical.
Reply with the single most
specific category.

model gpt-5.4 · temperature 0.2 · unchanged

Cases that flipped · pass → fail

"The checkout page is down and we are losing orders every minute."

urgenttechnical

JudgeMissed the urgency signal on a production outage with revenue loss.

"I was charged twice and my card closes tomorrow."

urgentbilling

JudgeDouble charge with deadline pressure should escalate as urgent.

"Can you send me the invoice for last month?"

billingThe message is about billing.

JudgeReturned prose instead of the required single lowercase category.

Recommendation: Block the release. Restore the escalation rule in v17, then rerun the suite.

Methodology

What this study does—and does not—show

This is a worked, demo-safe benchmark of the release-review method. It is not a claim about general model quality or a production customer result.

Controlled comparison

Only the prompt version changes between the baseline and candidate.

Visible failure contract

Urgency handling and a single lowercase category are explicit pass conditions.

Reviewable release artifact

The prompt diff, failed cases, judge reasons, and decision remain together.

Not a universal model score

The pass rates do not generalize beyond this demonstration case set and prompt.

Representative pass-to-fail cases

These cases reveal two separate regressions: the candidate underweights explicit urgency signals and stops respecting the exact output format.

The checkout page is down and we are losing orders every minute.

urgenttechnical

The candidate ignored the revenue-impact escalation rule.

I was charged twice and my card closes tomorrow.

urgentbilling

The narrower category displaced the explicit deadline escalation.

Can you send me the invoice for last month?

billingThe message is about billing.

The candidate broke the required one-label output contract.

Run the same release check on your prompt

Keep the model and cases fixed, compare the candidate with a known-good version, and share every failure behind the decision.