First genuinely live-data agent audits of three golden properties: what a consultant would question
House (10k, owner): LGT_LED+ENV_WALL selected, PV 5.16 kW financeable (SIR 1.36), 224 s, ~0.56 USD, 0 failed tool calls, 0 correction turns. Culver office (150k): one 44.7 kW PV array takes the whole budget (SIR 2.03), 274 s. Houston office (50k): empty plan, every measure SIR<1 (PV 0.25, LED 0.39), 238 s. Findings that unit tests could not show: live office models had no windows ('No windows were found', WIN_* and LAYOUT infeasible) because the vision/WWR path was not feeding windows; baselines said 'not benchmarked: no climate zone' so benchmark checks never ran live; newer Vexcel captures made roofs differ from the examples (191 vs 151 m2). Read each recording's RunResult.warnings and the skipped reasons, not just the plan, before calling a recording good.