Supplementary material · interactive

Theoros, Version A and Version B

The study compared an icon-based recommender with a guide in which a language model reads the visitor and explains its choices while code keeps time and facts. Step through both versions as participants saw them, side by side.

Same task in both versions · “You are at Syntagma Square and have 2 free hours”

Step through both versions

1 / 6

What differs between the versions

Select a row to see the matching screens above.

Version A · ACUX-RVersion B · Theoros
Version B on a phone · the published app after the study, signed in with Google

One plan, explained

The model chose and explained the stops; code computed every time and handled “Not for me”. Select a number to see who wrote each part.

Written by the modelComputed by code from data
The plan
After “Not for me” on the first stop

The reason for the Currency Museum repeats an error in our place description, the archaeologist’s first name (Heinrich Schliemann). The model copied its source faithfully; the description has since been corrected.

Also in the app

Outside the study, Theoros also has a guide chat about the history and monuments of Athens and Attica, a camera lens for places and audio. These features were switched off during both study versions, so that the comparison covers only the profile and the recommendations. Generative features need a study link or sign-in with Google; the app works without them.

How Version B makes a plan

The model does two jobs only: it reads the visitor into the eight ACUX profiles, and it chooses, orders and explains stops from a list that code has already filtered. Every clock time, opening hour and walking time comes from code and verified data.

Model (Gemini, locked prompt)Code and verified data
1 · profiler v1.1

Read the visitor

Photo likes give weights; one sentence is scored against the eight written profiles, with the evidence for each score.

2 · candidates

Filter places

Only places that can be visited in full, walk included, within opening hours and the budget. Outdoor places drop out in rain.

3 · planner v2.2

Choose and explain

Ordered stops, one reason per stop from its description, one exploration stop and explained alternatives. No clock times.

4 · scheduler

Keep time

Walking time and visit length give the times; stops that do not fit are dropped and gaps filled from the alternatives.

5 · during the walk

Adapt

Closing, running late, rain or heat trigger one proposal at a time. No new model call.

Why code keeps time

Six planner versions on five test scenarios, two runs each. While the model wrote the clock times, adding temporal rules to the prompt made plans worse. Moving scheduling to code made all ten plans valid.

All checks passed Feasible as proposed Locked for the study

v1.0 ran at a higher thinking level with fewer checks (median latency 36 s); v1.1 to v2.2 at low thinking level, 1.9–2.6 s.

The study

A within-subjects study: every participant did the same task with both versions, in an order the server alternates, and rated each version right after using it. Participation is anonymous; the app shows only an automatic code.

Consent, 18+›3 background questions›First version›UEQ-S + 6 items›Second version›UEQ-S + 6 items›Which version would you use?
8–17 October 2026 · 21 participants used both versions

What participants said

Both versions were rated highly. Version B was rated higher mainly in hedonic quality, the UEQ-S scale for stimulation and novelty, and more participants would use it on a trip. Most other ratings were at or near the top of the scale in both versions, so they could not differ much.

21participants completed both versions (10 in class, 11 remotely)
Excellentboth versions on the UEQ-S benchmark, on all three scales
+0.51hedonic quality, B over A (r = 0.63; p = 0.004, Holm 0.031)
12 : 5would use B : A on a next trip (4 no preference)
Show results for

All participants. Hover or focus a mark for its values.

UEQ-S against the benchmark

Version A Version BLine: 95% confidence interval. Bands: Hinderks et al. (2018)

Who rated which version higher

◀ A highersameB higher ▶
Results table: medians (IQR), Wilcoxon signed-rank test, all 21
MeasureVersion AVersion Bzp (Holm)r
UEQ-S pragmatic (−3 to +3)2.50 (1.50–3.00)2.75 (1.75–3.00)+1.071.0000.23
UEQ-S hedonic (−3 to +3)2.00 (1.25–3.00)2.25 (2.00–3.00)+2.880.0310.63
Perceived accuracy (1–5)4 (4–5)4 (4–5)+0.581.0000.13
Transparency (1–5)5 (4–5)5 (5–5)+1.311.0000.29
Trust (1–5)4 (3–5)4 (4–5)+0.261.0000.06
Confidence (1–5)4 (3–5)4 (3–5)+0.491.0000.11
Use intention (1–5)5 (4–5)5 (4–5)+0.821.0000.18
Profile accuracy (1–5)4 (3–5)4 (4–5)+0.771.0000.17

In their words

The class session took place in a course taught by the researcher, who was present; the advantage of Version B was larger in that group, though not significantly. Use the buttons above to compare.