EventProof

Before an AI camera fines anyone, EventProof tests whether the rule would hold up. Cities already issue tickets from AI cameras (Takoma Park, MD: $40 stop-sign citations; New York's AI bus-lane cameras issued about 3,800 mistaken violations in 2024 because the rule was wrong, not the detection). A stop law says "shall stop" but never for how long, or how slow counts as stopped, so every camera rule silently picks both numbers. Same law, same 30 real cars: anywhere from 0 to 17 tickets. On 1,268 real STOP approaches at US intersections, the verbatim California statute would ticket 9% to 85% of drivers, only by those two numbers. How it works: 1. Compile: Muse Spark turns a sentence or a verbatim statute into a typed rule (never code); every value is quoted from the text or marked as assumed. SAM 3.1 tracks every car; a deterministic engine judges each one with its frames. 2. Attack: every ticket is re-judged under small, reasonable changes. Muse Spark argues the driver's defense from a closed menu and the engine decides which arguments win; Muse Wonder reads the rule a second time. Of 20 real tickets, all correct under the rule, only 4 survive every challenge. 3. Withhold: while a ticket rests on a choice nobody approved, the rule issues nothing. Once a person signs the open numbers, 17 of 20 issue automatically (0 wrong) and 3 go to a human with the reason. Agentic loop with a deterministic referee: models compile, draft zones, argue the defense and propose repairs; code verifies every claim; a repair is kept only with zero regressions on hidden tests; a person signs what stays open. Every test is pre-registered before any model call and recomputed in CI (real footage from Serbia, Poland and Brazil with frozen human labels, a sealed synthetic holdout, 40 repeated statute compiles). Limits are stated in the README: the labelled real sets are small. Built with Muse Spark, Muse Wonder and SAM 3.1.