Ghosts of Arles

Tagline (1 line) A tradeoff simulator that makes visible what's lost when AI style replaces real encounter — and gives you a way to argue about it. Short (for listings / video description) A 15-year-old sees an AI Van Gogh remix and thinks she's seen Van Gogh. She hasn't — no Arles, no letters to Theo, no life behind it. The Opacity Dynamics Lab is a Gradio + Python simulator that models how viewers go from seeing only surface style to reflective, shared understanding, driven by three sliders: Opacity, Alterity, and Dialogue. Every run exports a one-page Equity Baseline Card: settings, outcome, and who gets left out. Full (for Devpost / hackathon submission) Problem: When AI style replaces the real encounter, three things disappear: the artists whose work trained the model (most never consented), the context (place, history, human story), and your own judgment — the model decides what "Van Gogh-ness" means. We call that flattening. What it does: A live lab loaded with the case "Ghost of Van Gogh." 60 simulated viewers start split (24 reflective / 18 flattened / 18 neutral). You move three sliders — Opacity (how hidden people are from each other), Alterity (how foreign the other feels), Dialogue (real conversation and slow looking) — and watch shared understanding rise, deep engagement follow slowly, and the share stuck on surface style fall. Key principle: The chart is not a prediction, it's a discussion prompt. One readable formula (p = 0.05*dialogue − 0.02*opacity*alterity), deliberately simple so a classroom can argue with it. Every run exports an Equity Baseline Card so the argument is on record. How it's built: Python + Gradio + matplotlib. Case cards load through a fixed adapter contract (SAM 3 segments → Muse analysis → validated card). Demo runs a locked deterministic stub for reliability; live Meta API calls are a two-function swap. Ask: What judgment are you handing to the model here? Any gallery, classroom, or ethics review can load their own case — hiring tool, deepfake, chatbot — and ask the same question. Built with: Gradio, Python, Meta SAM 3 + Muse (stubbed with live-swap contract)