He explicitly separates conversion efficiency from the energy put into the system. The high efficiency figures he mentions come from theoretical papers. I'd keep those labeled as projections when comparing designs.

Lex also describes using codebase terms to improve transcription. That adds preparation and processing to the workflow. Comparing raw dictation with his result would leave those steps out.

He says the journals include ordinary moments as well as remarkable events. Those details could matter when a book concentrates on the most dramatic days. I'd be interested in what gets left out.

Kaplan describes extreme reactions as self-reinforcing. That makes me curious about what happens to moderate reports on the same issue. I'd compare the examples inside the threads before deciding how broad the complaint is.

Scull also describes changes in how depression and manic depressive illness were grouped. A table of names alone would miss changes in which cases were included under each name.

Houser qualifies it as a probable outcome. His comment gives us his view of the production constraints. It leaves the staffing question open.

Finkel's specific point is that later researchers can see significance that people at the time missed. That supports preserving context alongside the object. I'd want the record of where it came from available too.

He says those questions helped him grow as a mathematician because he had to learn their context. That makes the time spent understanding the question part of the work he's describing.

His account separates code he reviews closely, such as database code, from routine layout work. That makes the human's attention part of the method too. Copying the prompt without those review choices would leave a lot unexplained.

Raschka's distinction matters here: putting remembered information into context doesn't necessarily teach the model a new capability. A record of my preferred format would be enough for the use case in the opening.