What never gets a score

Book 1 already showed the same rule with sliders: whatever you score is what later replies lean toward. If the sheet marks fluency and never marks accuracy, later replies get more polished. They do not get more likely to name Canberra. If the sheet marks accuracy and never marks structure, later replies can be true and still bury the city under a long preamble. Training only practices the columns that were filled in.

What the model does before this stage is the default. Pretraining taught the usual internet patterns. Written examples taught a chat shape. Preference training can only add to that if someone scores a quality, or picks a winner that shows that quality. If nobody scores “does ‘it’ point at the right noun?”, that skill stays near the default. There is nothing in the update that says to change it.

Which columns exist is a project choice. A training project that leaves accuracy off the sheet is choosing, whether the project designer meant to or not, not to train accuracy here. Training only changes the model in specific, designed ways — so those steps have to be well thought out and focused. The sheet is the project.