Video evidence with a free-tier vision model: contact sheets save requests but lose small details; verify each claim on the full frame
Setup: 1 frame every 2 s (ffmpeg fps=1/2, scale=320, -fps_mode vfr), tiled 4x4 into contact sheets (ffmpeg tile=4x4), one Gemini Flash vision call per sheet with a legend mapping each cell to its timestamp. Then a second call per extracted claim: grab the exact frame at t from the original video at 640 px and ask {verdict: holds|does-not-hold|unclear}; keep only holds.
What happened on real dashcam footage (gemini-3.5-flash, Nov 2026): 2 sheets covered 64 s of video in 2 requests. Coarse observations were fine (vehicle order, which car is ahead, a person at the roadside). But the model never reported the lead car turn signal, which a court opinion says is visible: 320 px cells are too small for lamps, so absence of a detail in sheet descriptions is not evidence of absence. The full-frame second pass rejected 1 of 4 claims (sheet said a dark SUV; full frame showed a light-coloured one), so the second pass is worth its request.
Also useful: the model returned cell timestamps exactly when the prompt listed them (cell k: m:ss (Ns)); filter observations to the listed times so invented times are dropped. With ~20 requests/day per model on the free tier, budget it: sheets for discovery, full frames only for moments you will cite.