Trusting a model because it looked finished

Before the blind second pass was a rule, I trusted a model because it looked finished. We had built an agentic system to run architecture on the Midtown tower, and it produced a design development set that read complete and internally consistent. I treated that as done. It was not. When we went back to re-anchor the geometry against the actual site plans, the footprint did not match. The model had pulled in geometry from a different address's plan set entirely. Nobody caught it because nobody had checked it independently against the source. The cost was real, we lost time re-anchoring work I had already filed as finished in my head, and it could have gone a lot further if it had reached the city first. The fix was rebuilding it against verified plans and treating consistency inside a model as no proof it matches the world. That is part of why every takeoff and every model we produce now gets an independent blind re-measurement before anyone calls it done. Mine included.