evacuee
step 1first, a model of the town.
it matches the real numbers — but it has no idea who anyone is.
an illustrative evacuation run — real Seaside / Cascadia tsunami data (not Nanao). agents flee by simple rules.
tuned to fit the real Nanao surveyevacuated —%by car —%error —
turn the dials.
how readily people evacuate2.60
how often evacuees take the car62%
how often they head to high ground42%
does a go-bag make you leave? (− = less)-1.20
outcome vs the real Nanao survey
evacuated87% · 88%
by car63% · 62%
to high ground40% · 42%
go-bag owners' odds0.34 · 0.30
tick = real Nanao · green = match
data & ground truth: Hasegawa, Asai, Yamaguchi, Inada, Yamazaki, Takemoto, Takabatake (2025), “Tsunami awareness, preparedness, and evacuation behaviors during the 2024 Noto Peninsula Earthquake Tsunami”, Safety Science 189 · N=232 · doi:10.1016/j.ssci.2025.106885
● real run — the decision & transcript pages and “what survives” use a real LLM simulation on the Noto road graph. ○ illustrative — “which explanation”, “the loop”, “your turn” and the sandbox are interactive client-side toys for intuition, not results. The Seaside animation is illustrative (not Nanao); model “did not reach safety” counts are hypothetical, not real Nanao outcomes.
● real run — the decision & transcript pages and “what survives” use a real LLM simulation on the Noto road graph. ○ illustrative — “which explanation”, “the loop”, “your turn” and the sandbox are interactive client-side toys for intuition, not results. The Seaside animation is illustrative (not Nanao); model “did not reach safety” counts are hypothetical, not real Nanao outcomes.
