Attach none and groundedness is not scored — the composite is the weighted mean of whichever factors apply.

Equal weights average to a plain arithmetic mean. Scores run 0–100.

Reserved for a later version. Human review applies to ad-hoc use, not to a controlled study, because a per-enrollee edit is not a uniform condition.

How much is remembered is a platform setting, not a study setting — it is capped by the smallest context window among the models in the set-up, so every arm of the comparison gets the same. Enrollees can rewind or clear a session themselves.

One artifact holding both the questioning and the rewrite phrasings. Required once this study attaches context documents or declares required details — the assistant would otherwise have nothing to reply with. Choose the style per run at launch, to measure which gets people to an answerable request faster.