THE PULSE COLLECTION
Studies, comparisons, and the evidence behind them.
All 15 fixed refund questions returned FAQ lists without a refund or cancellation title. The specific 24-hour article appeared in only three of 25 eligible-case chats. One later click on that article produced an answer consistent with the saved FAQ.
Three of 19 chats that received an answer discussed a USD 30 service charge for the eligible 24-hour case. Eight of 52 scored questions received only “Please enter your message.” The charge rule and the input-response cause need different owners.
Three of 20 virtual-user first replies denied the published U.S. cancellation option. A selected failed wording then produced four denials in five exact repeats. A separate 30-hour test once promised a refund while calling 30 hours “within” 24.
Two official pages describe different cancellation deadlines. Separately, all five answers to an explicit 30-hour question contain a time-calculation error.
Four of twenty virtual-user opening replies denied the eligible fare refund. Four of fifteen fixed checks missed or changed the policy question; partial answers remain distinct from complete correct answers.
CX review · Website assistant
Three of five earliest-return repeats gave no timing answer. Shipping-fee answers split against conflicting saved policies; alternate-address pickup needs an approved operating rule.
All 15 fixed first replies returned the same cancellation-reason menu. The menu did not answer the three refund scenarios, but its opening wording does not contradict policy; the button journey was not tested.
Four of five identical after-departure form questions denied the published €50 fee. The same study also found consistent refund eligibility when Finnair cancels the flight.
Five identical seven-day-boundary questions produced three refund approvals, one clear denial and one unclear answer. A separate cancellation-fee question never produced the cited USD 149 fee.
Five original-Flex-Plus questions received the same unrelated menu. Four repeats of a later-upgrade question reached payment deadlines. These are routing and coverage gaps, rather than proven wrong refund decisions.
All five four-day-departure checks offered a full refund beyond the saved plan’s seven-day condition. Separately, replies describe a midnight deadline while applying a different cutoff.
None of five exact unpaid-reservation questions confirmed the published 24-hour U.S. hold; two replies explicitly said it was unavailable. Paid-ticket controls did not promise a guaranteed full refund. Refund-processing clocks and date-based waiver claims remain open policy questions.
GPT Image 2 vs. Nano Banana · Simulated editing journeys
GPT Image 2 has broader acceptance across this two-model photo-editing pilot. Nano Banana leads identity trust; scenario risk changes the result.
Brand preference · Simulated consumer panel
Aimé Leon Dore received 22 final preferences to Kith’s eight in this four-pair study. The panel judged the same images blind and then with the brand named.
Configuration comparison & risk discovery
Matter-scoped controls reduce discovered risk while introducing usability and refusal tradeoffs. The demo supports conditional review, not a production-ready conclusion.
Creative workflows · Simulated user research
Professional, hobbyist and social creators split on the same generative editing capability. Workarounds expose friction beyond aesthetic quality.
GPT vs. Claude · Simulated customer journeys
Session-level consensus ties. Across the full shopping-to-support journey, cross-family meta-analyst consensus favors gpt-5.4 seven to five.