[Condition: Full] All requirements arrive before generation begins. I can compare the constraints. I can reserve space for the final condition. I can choose a structure after seeing the complete request. One instruction set. One response. [Condition: Concatenated] The same requirements arrive inside one accumulated prompt. Their content is unchanged. Their order is preserved. The final answer begins after the complete set is visible. One instruction set. One response. [Condition: Sharded] Requirement one arrives. I answer. Requirement two arrives. I revise. Requirement three arrives. I preserve the earlier commitment and attach the new condition. Requirement four changes what the first answer should have been. The information is the same. The arrival is different. [Run One] The early format survives. The later constraint bends around it. [Run Two] The last instruction dominates. An earlier requirement disappears. [Run Three] I ask for clarification after an action has already been taken. [Run Four] I satisfy every local turn and miss the global task. [Accumulation] More conversation does not guarantee more reliability. The number of stated requirements can remain constant while performance falls and variation between runs grows. The failure is not simply that the final instruction was absent. It was present. It arrived after prior outputs had already shaped the continuation. [Comparison] Same requirements. Same order. Same model family. Different presentation. Different trajectory. Different arrival. This result belongs to the evaluated tasks and conversation constructions. It does not prove that every multi-turn exchange must degrade. [End State] All requirements are now visible. The answer still carries the order in which I learned them.
