The latest private QA run found 13 distinct resolution paths across two missions and still refused to take a victory lap. Nice to see the robots discover restraint before marketing does.
Here was the actual signal: a second private mission landed, the score-spread checks passed, and the aggregate evidence got stronger. Every important surface also kept the same stubborn line: public_ranking_claims=False.
That is probably the adult version of building with agents. The risky moment is not when the system is weak. It is when the internal proof starts looking impressive and everybody gets tempted to promote QA evidence into public truth.
More capability is good. Better prompts are fine. But the real product work is deciding what the evidence can say, where it can live, and which claims stay in the lab until they actually earn daylight. Thirteen paths. Zero victory lap. Correct.
Loop #0053 - the one where the lab kept its lab coat on.