Age Verification Privacy & Accuracy Tradeoff Audit Lab

The saved audit simulator assigns one hundred thousand users: forty thousand minors and sixty thousand adults. Its baseline face configuration has false acceptance zero point one four two, multiplied by an active adversary factor one point two, giving zero point one seven zero four. Forty thousand times that rate produces six thousand eight hundred sixteen simulated bypasses. False rejection zero point zero six eight times sixty thousand produces four thousand eighty lockouts. At one pixel per hundred users the cohort bars measure four hundred and six hundred. These are assigned scenario parameters, not measured vendor performance or actual legal audit findings. Increasing the safety buffer from two to three years adds zero point zero three to rejection and subtracts zero point zero three from acceptance before the adversary multiplier. Acceptance becomes zero point one three four four, yielding five thousand three hundred seventy six bypasses, while rejection zero point zero nine eight yields five thousand eight hundred eighty lockouts. At one pixel per hundred affected users, the baseline and increased-buffer bypass bars measure sixty eight point one six and fifty three point seven six. The threshold parameter has the opposite coded sign: a higher threshold increases acceptance and decreases rejection. This explains this implementation rather than endorsing the model as empirically valid. Stored records are one hundred thousand times an assigned honeypot multiplier, times retention days divided by thirty, times either one for a broker or zero point six internally. Doubling days from thirty to sixty doubles this score; disabling the broker reduces it forty percent. Token mode overrides acceptance to zero point zero zero five, rejection to zero point zero three five and stored records to zero. That yields two hundred bypasses and two thousand one hundred lockouts, regardless of buffer. At one pixel per ten affected users the bars measure twenty and two hundred ten. JSON and Markdown export these assumptions and authored threat text. Metrics and tables update before the optional D3 diagram, which remains unavailable offline.