TableFairDoes fair picking help real groups?
Two kinds of evidence: a controlled evaluation on live Qloo data, and live, anonymous numbers from groups using TableFair.
Held-out test against Claude alone
57 random groups with real constraints (a time, a budget, a vegetarian, children, a time limit). Each person had four favourites; TableFair and Claude saw three, and every pick was graded on the fourth. Method
Least-matched, held-out
41 vs 18
percentile: TableFair vs Claude alone (no Qloo)
Claude alone breaks a rule
27%
closed, over budget, not vegetarian, wrong age rating or too long
TableFair better
34 of 47
groups vs Claude alone (worse in 11)
vs simple average
Draw
40.6 vs 40.1 on held-back tastes
vs most popular
41 vs 34
films 37 vs 13; dinner 44 vs 49, popular ahead
Evaluation on live Qloo data
120 random groups of 3–5 people (60 dinner, 60 movie) in two runs, same pipeline as the app, no LLM. Synthetic groups. Method
Fair pick ≠ average
42%
50 of 120 groups
Lift for least-matched
+13.5
points, where they differ
Cost to group average
−6.8
points, where they differ
Flexible tonight
37%
of people: Qloo saw little difference for them
Without taste data
52%
of groups leave someone below 25% (fair pick: 0%)
Least-matched, no taste data
24%
vs 63% with the fair pick