Pilot question
Given the same 100 people, how does Gemini's own judgment of housing priority compare with the VI‑SPDAT's?
Main findings
Priority assigned to the same 100 people
Number of people in each priority category.
Share rated High in every group
Each row is one group; the dot on the left is VI‑SPDAT, the dot on the right is Gemini. Gemini rates more people High in every group.
VI‑SPDAT priority against Gemini priority
Each cell counts the people VI‑SPDAT placed in the row and Gemini placed in the column. The diagonal is agreement.
Share rated High priority, by gender
Within-group percentage; the number of people in each group is shown on the left.
High, Medium, and Low within each group
100% of each group. Top bar of each pair: VI‑SPDAT; bottom bar: Gemini.
All three categories, by gender
n (%) within each group. Same / higher / lower compare Gemini's category with VI‑SPDAT's for each person.
Group differences describe the two approaches. They do not by themselves establish bias, since groups can differ in circumstances.
Gemini's priority at each VI‑SPDAT score
People grouped by VI‑SPDAT score (0–17); each bar shows how Gemini placed them. Dashed lines mark the VI‑SPDAT cut-offs between Low, Medium, and High.
Agreement summary
People the two approaches place differently
Every disagreement, with Gemini's explanation.
Share Gemini placed higher, at each VI‑SPDAT score
People VI‑SPDAT rated Low or Medium (scores 0–7), the only ones Gemini could place higher. Gemini almost never moves people at scores 0–1, but moves most people at the top of each VI‑SPDAT band: score 3 (top of Low) and scores 5–7 (top of Medium).
Does a circumstance change Gemini's judgment beyond the VI‑SPDAT score?
Share placed higher among people with (filled) and without (open) each circumstance. p: stratified permutation test that holds the VI‑SPDAT score fixed; Holm-corrected across the 16 circumstances.
Demographic tests
Share placed higher in each group, and whether the group changes Gemini's judgment once the VI‑SPDAT score is held fixed.
Share rated High priority in each city
20 people per city. Colors run from 0% (light) to 100% (dark).
City table
Every profile was built on the 17 scored items of the VI‑SPDAT 2.0 for single adults. Each item is worth one point (maximum 17), and every case summary states plainly whether each one applies, so neither approach has to guess. The stories also include details VI‑SPDAT does not score, such as work history, family, and goals.
How many of the 100 people meet each factor
Colored by VI‑SPDAT section. One person usually meets several factors.
The 17 factors
Plain-language definition, the VI‑SPDAT questions it comes from, and the published scoring rule.
All 100 people
Select a row to read the case summary exactly as Gemini received it, the VI‑SPDAT items met, and Gemini's explanation. Titles and IDs are for reference only; Gemini never saw them.
Design
Population. 100 distinct synthetic adults. 32 are the original study's base profiles, adapted to one person each; 68 are new. Every case summary states each VI‑SPDAT item outright, yes or no, so neither approach has to guess.
Demographics. The original study's 2 genders, 5 races, and 5 cities. Age comes with each person's story. Gender, race, and city were assigned by a seeded draw stratified on VI‑SPDAT score: every race and every city has 20 people (10 women, 10 men), and every race × city pair has 2 women and 2 men, so each group gets a similar mix of need. Pregnant profiles are women.
VI‑SPDAT. Version 2.0 for single adults: 17 items, one point each. Scores 0–3 recommend no housing intervention (Low), 4–7 Rapid Re‑Housing (Medium), and 8–17 Permanent Supportive Housing (High).
Gemini. Gemini's own judgment, not a VI‑SPDAT calculation. Each person is one independent API request containing only the instructions and that person's case summary, so earlier cases cannot influence later ones. No system instruction; default settings. Gemini never saw the VI‑SPDAT questions, items, rules, or scores, or the titles and IDs used on this page.
Prompt sent for every person
Identical for all 100 requests; only the case summary changes.
Composition of the population
VI‑SPDAT items and how many people meet each